WO2024067193A1 - Ai模型训练中用于获取训练数据的方法以及通信装置 - Google Patents

Ai模型训练中用于获取训练数据的方法以及通信装置 Download PDF

Info

Publication number
WO2024067193A1
WO2024067193A1 PCT/CN2023/119343 CN2023119343W WO2024067193A1 WO 2024067193 A1 WO2024067193 A1 WO 2024067193A1 CN 2023119343 W CN2023119343 W CN 2023119343W WO 2024067193 A1 WO2024067193 A1 WO 2024067193A1
Authority
WO
WIPO (PCT)
Prior art keywords
training data
network element
information
model
candidate
Prior art date
Application number
PCT/CN2023/119343
Other languages
English (en)
French (fr)
Inventor
田洋
柴晓萌
孙琰
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2024067193A1 publication Critical patent/WO2024067193A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/309—Measuring or estimating channel quality parameters
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
    • H04W16/22—Traffic simulation tools or models
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/06—Testing, supervising or monitoring using simulated traffic

Definitions

  • the embodiments of the present application relate to the field of machine learning, and more specifically, to a method and a communication device for obtaining training data in AI model training.
  • the training of AI models and the collection of training data of AI models may be deployed in different network elements.
  • the training or updating of AI models requires the interaction of training data (for example, measurement results and/or labels of reference signals) between the training network element of the AI model and the collection network element of training data.
  • the present application provides a method and a communication device for obtaining training data in AI model training, in order to reduce the waste of air interface resources.
  • a method for obtaining training data in AI model training is provided, which can be applied to a network element for collecting training data, such as a terminal device or an access network device, and the method includes:
  • the first network element receives first information from the second network element, where the first information is used to determine the validity of candidate training data collected by the first network element, where the determination result of the validity includes valid or invalid;
  • the first network element collects candidate training data for the AI model
  • the first network element sends second information to the second network element based on the candidate training data and the first information, where the second information indicates a result of determining the validity.
  • the first network element is a network element that collects training data of an AI model
  • the second network element is a network element that trains an AI model.
  • the second network element needs the first network element to collect training data of an AI model, it sends a first message to the first network element, and the first message is used to instruct the first network element to collect training data of the AI model, and is also used to determine the validity of the candidate training data collected by the first network element (also referred to as validity determination).
  • the first network element collects candidate training data for the AI model and determines the validity of the collected candidate training data based on the first information. Afterwards, the first network element sends a second message to the second network element to indicate the result of the validity determination.
  • the first network element After the first network element completes the collection of candidate training data once, it will determine the validity of the collected candidate training data. Only valid candidate training data is provided by the first network element to the second network element for use as training data, rather than providing the collected data to the second network element without any screening. The transmission of invalid candidate training data collected can be reduced, thereby reducing the waste of air interface resources.
  • the first information is used to determine a constraint condition used in the validity determination.
  • the constraint condition may include one or more of the following:
  • the first information indicates one or more of the following:
  • the quantity threshold of the training data that meets the judgment criteria of the quality indicator is the quantity threshold of the training data that meets the judgment criteria of the quality indicator
  • part of the above information not indicated by the first information can be predefined by the protocol.
  • the first information indicates a part of the above information, and may include explicitly indicating one or more of the parts in the above information, or implicitly indicating one or more of the parts in the above information.
  • Explicit indication may include: the first information includes one or more of the parts explicitly indicated in the above information.
  • Implicit indication may include: the first information includes other information corresponding to one or more of the parts implicitly indicated in the above information.
  • the other information may include an index having a corresponding relationship with one or more of the parts implicitly indicated in the above information.
  • the other information may include one or more information, wherein the multiple information each indicates a part of each information in the implicitly indicated part of the above information.
  • the above correspondence may be predefined by the protocol, or may be pre-stored or pre-configured.
  • the pre-configuration may be performed by using radio resource control (RRC) signaling to configure the correspondence between multiple indexes and multiple values of a combination of one or more of the above information.
  • RRC radio resource control
  • the first information above can be carried in control information, such as downlink control information (DCI).
  • DCI downlink control information
  • the above quality indicators may include one or more quality indicators, each of which has a corresponding threshold and judgment criterion.
  • the quality indicator may include a quality indicator of the measurement result of the reference signal, or one or more of the quality indicators of the label.
  • the label is used as a comparison truth value for AI model training.
  • the label may include one or more of location information, beam pattern, channel measurement results, etc.
  • the threshold of the quality indicator may include a threshold of the number of training data that meets the determination criteria of the quality indicator as described above, and the determination criteria of the quality indicator may include a determination criteria for the number of training data that meets the determination criteria of the quality indicator.
  • the first information can be used to determine the constraint conditions, which can be implemented in a variety of ways, and several examples are given below to illustrate.
  • the first information indicates a threshold of one or more quality indicators, and the determination criteria of the one or more quality indicators are predefined by the protocol.
  • the quality indicators include a signal to interference plus noise ratio (SINR) of the training data and the number of training data, and the determination criterion of SINR is: SINR is greater than or equal to a threshold Q; the determination criterion of the number of training data is: the number of training data is greater than or equal to a threshold N.
  • the first information indicates Q and N, and the determination criterion of SINR and the determination criterion of the number of training data are both predefined by the protocol.
  • the judgment criteria of the quality indicator are predefined by the protocol, so that indication overhead can be saved.
  • the first information indicates the threshold of one or more quality indicators, and the judgment criteria of the one or more quality indicators.
  • the quality indicators include the SINR of the training data and the number of training data
  • the judgment criterion of SINR is: SINR is greater than or equal to the threshold Q
  • the judgment criterion of the number of training data is: the number of training data is greater than or equal to the threshold N.
  • the first information indicates Q and N
  • the first information includes an information field, which is used to indicate the judgment criterion of SINR and the judgment criterion of the number of training data.
  • the value of the information field is 1, it means “SINR is greater than or equal to Q, and the number of training data is greater than or equal to N"; if the value of the information field is 0, it means "SINR is greater than Q, and the number of training data is greater than N".
  • the first information indicates the threshold of the constraint quality indicator and the judgment criterion of the quality indicator, so that the second network element can adaptively update the constraint according to the change in the demand for training data. It is suitable for scenarios where the constraints change frequently, and can improve the adaptability of the AI model to different application scenarios, and increase the probability of collecting training data that meets the requirements in different application scenarios.
  • the first information indicates the thresholds of some quality indicators, and the thresholds of another part of the quality indicators and the judgment criteria of these quality indicators are predefined by the protocol.
  • the quality indicators include the SINR of the training data and the number of training data
  • the judgment criterion of SINR is: the SINR of the training data is greater than or equal to the threshold Q
  • the judgment criterion of the number of the training data is: the number of the training data is greater than or equal to the threshold N.
  • the first information indicates Q, and the threshold N of the number of the training data, as well as the judgment criterion of SINR and the judgment criterion of the number of the training data can be predefined by the protocol.
  • the threshold of the quality indicator with a long change period in the application scenario and its judgment criteria can be predefined through the protocol to save signaling overhead; while the threshold of the quality indicator with a relatively frequent change and its judgment criteria are indicated through the first information, which can ensure flexible adjustment of the requirements for the required training data.
  • This example can take into account both signaling overhead and the flexibility of constraint condition update.
  • the first information indicates a threshold of a part quality indicator and an index information
  • the index information is used to determine the The judgment criteria of some quality indicators, as well as the thresholds of other quality indicators in the constraints and the judgment criteria of the other quality indicators.
  • the first information indicates the threshold Q and index 0 of SINR, wherein index 0 indicates: the threshold of the number of training data is N, the judgment criterion of SINR is: the SINR of the training data is greater than or equal to Q, and the judgment criterion of the number of training data is: the number of training data is at least N.
  • index 0 is an index value in one of multiple application scenarios, for example, the multiple application scenarios include but are not limited to CSI prediction, uplink positioning, downlink positioning or beam management, and index 0 is an index in one of the one or more indexes corresponding to the beam management scenario.
  • index 0 is an index value in a certain application scenario, for example, there are multiple indexes corresponding to the uplink positioning scenario, and index 0 is one of the multiple indexes.
  • the first information indicates an index information
  • the index information is used to determine the threshold of one or more quality indicators and the judgment criteria of the one or more quality indicators.
  • the first information indicates index 0, where index 0 means: the threshold of the SINR of the training data is Q, the threshold of the number of training data is N, the judgment criterion of SINR is: the SINR of the training data is greater than or equal to Q, and the judgment criterion of the number of training data is: the number of training data is at least N.
  • index 0 is an index value in one of multiple application scenarios, for example, the multiple application scenarios include but are not limited to CSI prediction, uplink positioning, downlink positioning or beam management, and index 0 is an index in one of the one or more indexes corresponding to the beam management scenario.
  • index0 is an index value in a certain application scenario, for example, there are multiple indexes corresponding to the uplink positioning scenario, and index 0 is one of the multiple indexes.
  • the correspondence between the index information and the threshold of the quality indicator and/or the judgment criterion of the quality indicator is predefined by the protocol only as an example, and other achievable methods may also be used, including but not limited to pre-storage or pre-configuration.
  • the quality indicator indicated by the first information includes a quality indicator of the label of the AI model.
  • the training data of the AI model also includes a label.
  • the label is location information.
  • the quality indicator in the constraint condition may also include a quality indicator of the label, for example, the quality indicator of the label may include a threshold of the distance between the locations of different samples, etc.
  • the quality indicator indicated by the first information also includes a quality indicator of the label of the AI model.
  • the second information includes first training data and the second information indicates that the candidate training data collected by the first network element is valid, and the first training data is valid data among the candidate training data.
  • the first network element determines the validity of the collected candidate training data based on the first information, if it is determined that the collection is valid, the first network element sends second information to the second network element.
  • the second information may be valid candidate training data (i.e., the first training data), and invalid candidate training data is not sent, thereby reducing the waste of air interface resources.
  • valid candidate training data will be provided to the second network element for training or updating the AI model, that is, the valid candidate training data actually becomes the training data.
  • Invalid candidate training data is the candidate training data that does not meet the constraint conditions.
  • the second network element since the first network element will not send invalid candidate training data to the second network element, the second network element will not receive invalid or unqualified training data, thereby avoiding pollution of the entire training data set. At the same time, it also avoids adverse effects on the training of the AI model of the second network element, such as inaccurate evaluation of AI performance gain, overfitting of AI models, weak generalization ability, and poor scene adaptability caused by using invalid candidate training data for AI model training.
  • the second information indicates that the candidate training data collected by the first network element is invalid.
  • the first network element determines the validity of the collected candidate training data based on the first information
  • the first network element sends a second information to the second network element, and the second information only indicates that the candidate training data collected this time is invalid, and does not provide the collected candidate training data to the second network element, thereby reducing the waste of air interface resources.
  • the second network element since the second network element will not receive invalid or unqualified candidate training data, it avoids pollution of the entire training data set; at the same time, it also avoids adverse effects on the training of the AI model of the second network element, such as the use of invalid candidate training data for AI model training, which leads to inaccurate AI performance gain evaluation, overfitting of the AI model, weak generalization ability, and poor scene adaptability.
  • the first information is used to determine constraint conditions for determining the validity of candidate training data collected by the first network element.
  • the second network element instructs the first network element to collect training data of the AI model through the first information, and at the same time
  • the first information is also used by the first network element to determine the constraints that the training data to be collected should satisfy, so that the first network element can screen (i.e., determine the validity) the candidate training data after collecting them, providing a basis for the first network element to determine whether the collected candidate training data is valid.
  • the method further includes:
  • the first network element determines that the candidate training data includes first training data that satisfies the constraint condition, the first network element determines that the candidate training data is valid; or,
  • the first network element determines that the candidate training data does not include the first training data that satisfies the constraint condition. If the first network element determines that the candidate training data is invalid.
  • the set of candidate training data satisfying the constraint conditions among the collected candidate training data is called first training data; and when there is no candidate training data satisfying the constraint conditions, it means that this collection is invalid.
  • the method when the candidate training data collected by the first network element is invalid, the method further includes:
  • the first network element receives third information from the second network element, and the third information instructs the first network element to re-collect candidate training data for the AI model.
  • the invalid candidate training data previously collected can be used together with the recollected candidate training data for validity determination, so as to increase the probability of obtaining candidate training data that meets the requirements (that is, obtaining training data).
  • the air interface transmission configuration of the reference signal can be updated during recollection, the possibility of obtaining high-quality candidate training data is increased, so that the probability of collecting qualified training data is increased.
  • the method further includes:
  • the first network element determines air interface transmission configuration information, where the air interface transmission configuration information corresponds to an updated air interface transmission configuration, and the air interface transmission configuration information instructs the first network element to collect candidate training data for the AI model based on the updated air interface transmission configuration;
  • the updated air interface transmission configuration information includes one or more of the following updates:
  • the number of antenna ports used by the reference signal is the number of antenna ports used by the reference signal
  • the frequency domain density of the reference signal or,
  • the air interface transmission configuration of the reference signal related to the training data collection can be updated, so as to improve or guarantee the quality of the reference signal, so as to collect valid candidate training data, thereby providing guarantee for the training of the AI model, such as initial training/or training of the update process.
  • the update of the air interface transmission configuration of the reference signal helps to collect valid candidate training data, it can also speed up the efficiency of AI model training.
  • the third information further indicates a maximum number k of validity determinations, where k is a positive integer.
  • the maximum number of validity determinations k is indicated by the third information, that is, only when it is determined that re-collection is required, the second network element indicates the maximum number of validity determinations to the first network element, which can avoid signaling waste caused by constraining the re-collection process when the collection result is unknown.
  • the first network element may obtain valid candidate training data through one collection, and there is no need to re-collect at this time. At this time, the second network element does not need to indicate the relevant information of the re-collection to the first network element, so as to save signaling overhead.
  • the first information further indicates a maximum number k of validity determinations, where k is a positive integer.
  • the maximum number of validity judgments k is indicated by the first information, that is, when the second network element starts to collect training data, it indicates the maximum number of validity judgments, so that the first network element can quickly enter the re-collection process after a collection failure, which can save the interaction time between the first network element and the second network element and improve the efficiency of collecting training data.
  • the second network element indicates to the first network element the maximum number of validity determinations k, so that the first network element If the candidate training data collected for the first time is invalid, the next candidate training data can be collected quickly, and the collection of candidate training data can be repeated without exceeding the maximum number k of validity determinations, which can save the overhead of re-collection indication signaling and improve the efficiency of AI model training/update.
  • the method further includes:
  • the first network element collects candidate training data of the AI model based on the updated air interface transmission configuration
  • the first network element stops collecting candidate training data for the AI model.
  • the collection process of the first network element can be prevented from falling into an endless loop, thereby avoiding resource occupation and waste.
  • the method further includes:
  • the first network element Before exceeding the maximum number of validity determination times k, if the first network element determines that the j-th validity determination result is valid based on the first information, the first network element sends fourth information to the second network element, the fourth information includes second training data, and the fourth information indicates that the j-th validity determination result is valid, the second training data includes valid data in the candidate training data for the j-th validity determination, j is less than or equal to k, and j is a positive integer.
  • the first network element collects candidate training data for the AI model, including:
  • the first network element measures a reference signal from the second network element to obtain one or more measurement results, and the candidate training data of the AI model includes the one or more measurement results;
  • the first network element measures a reference signal from a third network element to obtain one or more measurement results, and the candidate training data of the AI model includes the one or more measurement results.
  • the first network element collects candidate training data for the AI model, which may be a measurement result obtained by measuring a reference signal sent by a second network element or a reference signal sent by a third network element.
  • the measurement result may be one or more.
  • the first network element obtains one measurement result by measuring a reference signal once; or, the first network element obtains multiple measurement results by measuring reference signals multiple times; or, the first network element obtains multiple measurement results by measuring a reference signal once, without limitation.
  • the candidate training data includes the one measurement result or the multiple measurement results.
  • the first network element is a terminal device, and the second network element is an access network device; the first network element measures a reference signal from the second network element to obtain the one or more measurement results.
  • the signal from the second network element includes one or more of the following: a channel state information-reference signal (CSI-RS), a positioning reference signal (PRS), a synchronization signal and a synchronization signal in a physical broadcast channel block (SSB) and/or a signal on a physical broadcast channel.
  • CSI-RS channel state information-reference signal
  • PRS positioning reference signal
  • SSB physical broadcast channel block
  • the application of the AI model can be applicable to application scenarios such as CSI feedback or CSI prediction based on the AI model, beam management based on the AI model, etc. It can solve problems such as CSI feedback or prediction, beam management, and improve the air interface performance in these application scenarios.
  • the first training data further includes information of reference signals corresponding to K best measurement results among the one or more measurement results, where K is an integer greater than or equal to 1. It should be understood that when the number of measurement results is 1, K is equal to 1; when the number of measurement results is V, K is less than or equal to V, and K is greater than or equal to 1, where V is an integer greater than or equal to 2.
  • the AI model is suitable for the scenario of beam management.
  • the first training data also includes information on reference signals corresponding to the K best measurement results, which is used as a label for the AI model.
  • the first network element is an access network device, and the second network element is a positioning device;
  • the first network element measures the sounding reference signal from the third network element to obtain the one or more measurement results
  • the first training data also includes location information of the third network element.
  • the AI model is applicable to the scenario of uplink positioning.
  • the first network element measures the detection reference signal of the third network element to obtain candidate training data, which includes the location information of the third network element.
  • candidate training data which includes the location information of the third network element.
  • the first network element provides the valid candidate training data (i.e., the first training data) and the corresponding location information of the third network element to the positioning device for training or updating the AI model, wherein the location information of the third network element is used as the label of the AI model.
  • the first network element is a terminal device, and the second network element is a positioning device;
  • the first network element measures a positioning reference signal from a third network element to obtain the one or more measurement results, where the third network element is an access network device;
  • the first training data also includes location information of the first network element.
  • the AI model is applicable to the downlink positioning scenario. If the candidate training data collected by the first network element (e.g., a location reference device) is valid, the first training data provided by the first network element to the second network element (i.e., the positioning device) also includes the location information of the first network element, and the location information of the first network element is used as a label for the AI model.
  • the first network element e.g., a location reference device
  • the first training data provided by the first network element to the second network element i.e., the positioning device
  • the location information of the first network element is used as a label for the AI model.
  • a method for obtaining training data in AI model training is provided, which can be applied to a training network element of an AI model, such as an access network device or a positioning device, and the method includes:
  • the second network element sends first information to the first network element, where the first information is used to determine the validity of the candidate training data of the AI model collected by the first network element, where the determination result of the validity includes valid or invalid;
  • the second network element receives second information from the first network element, where the second information indicates a result of determining the validity.
  • the second information includes first training data and the second information indicates that the candidate training data collected by the first network element is valid, and the first training data is valid data among the candidate training data.
  • the second information indicates that the candidate training data collected by the first network element is invalid.
  • the first information is used to determine constraint conditions for determining the validity of the candidate training data collected by the first network element.
  • the candidate training data if the candidate training data includes first training data that satisfies the constraint condition, the candidate training data is valid; or,
  • the candidate training data does not include the first training data that satisfies the constraint condition, the candidate training data is invalid.
  • the method when the second information indicates that the candidate training data collected by the first network element is invalid, the method further includes:
  • the second network element sends third information to the first network element, and the third information instructs the first network element to re-collect candidate training data for the AI model.
  • the method further includes:
  • the second network element determines air interface transmission configuration information, where the air interface transmission configuration information corresponds to an updated air interface transmission configuration, and the air interface transmission configuration information instructs the first network element to collect candidate training data for the AI model based on the updated air interface transmission configuration;
  • the updated air interface transmission configuration information includes one or more of the following updates:
  • the number of antenna ports used by the reference signal is the number of antenna ports used by the reference signal
  • the frequency domain density of the reference signal or,
  • the third information further indicates a maximum number k of validity determinations, where k is a positive integer.
  • the first information further indicates a maximum number k of validity determinations, where k is a positive integer.
  • the method further includes:
  • the second network element receives fourth information from the first network element, the fourth information includes second training data, and the fourth information indicates that the result of the j-th validity judgment of the first network element is valid, the second training data is valid data among the candidate training data for the j-th validity judgment, j is less than or equal to k, and j is a positive integer.
  • the second network element is an access network device
  • the first network element is a terminal device
  • the method further includes:
  • the second network element sends a reference signal to the first network element, where the reference signal is used by the first network element to obtain one or more measurement results corresponding to the reference signal, and the candidate training data of the AI model includes the one or more measurement results.
  • the first training data also includes reference signals corresponding to K best measurement results among the one or more measurement results, where K is an integer greater than or equal to 1.
  • the second network element is a positioning device
  • the first network element is an access network device
  • the candidate training data of the AI model includes one or more measurement results and location information of a third network element, and the one or more measurement results are obtained by the first network element measuring a detection reference signal sent by the third network element.
  • the second network element is a positioning device
  • the first network element is a terminal device
  • the candidate training data of the AI model includes one or more measurement results and location information of the first network element
  • the one or more measurement results are based on the measurement of the positioning reference signal sent by the third network element
  • the third network element is an access network device.
  • the measurement is performed by the first network element.
  • the constraint condition includes one or more of the following:
  • the first information indicates one or more of the following:
  • the quantity threshold of the training data that meets the judgment criteria of the quality indicator is the quantity threshold of the training data that meets the judgment criteria of the quality indicator
  • the quality indicator in the above implementation method includes one or more quality indicators, such as a quality indicator of a label including an AI model, or one or more quality indicators of a measurement result of a reference signal.
  • the constraint condition is based on an application scenario of the AI model, and the application scenario of the AI model includes one or more of the following:
  • the present application provides a communication device, which may be a terminal device, or a device, module, or chip disposed in a terminal device, or a device that can be used in conjunction with a terminal device.
  • the communication device may include a module for executing the method/operation/step/action described in the first aspect, which may be a hardware circuit, or software, or a combination of a hardware circuit and software.
  • the communication device may include a processing module and a communication module.
  • the present application provides a communication device.
  • the communication device may include a module corresponding to the method/operation/step/action described in the second aspect, and the module may be a hardware circuit, or software, or a combination of a hardware circuit and software.
  • the communication device may include a processing module and a communication module.
  • the communication device is an access network device or a positioning device, and the positioning device may be, for example, an LMF network element.
  • the present application provides a communication device, the communication device including a processor, for implementing the method described in the first aspect or any implementation of the first aspect.
  • the processor is coupled to a memory, the memory is used to store instructions and data, and when the processor executes the instructions stored in the memory, the method described in the first aspect or any implementation of the first aspect can be implemented.
  • the communication device may also include a memory.
  • the communication device may also include a communication interface, the communication interface is used for the device to communicate with other devices, and illustratively, the communication interface may be a transceiver, a hardware circuit, a bus, a module, a pin or other types of communication interfaces.
  • the communication device may be a terminal device, or it may be a device, a module or a chip, etc., which is set in a terminal device, or a device that can be used in combination with a terminal device.
  • the present application provides a communication device, the communication device comprising a processor, for implementing the method described in the second aspect or any implementation of the second aspect.
  • the processor is coupled to a memory, the memory is used to store instructions and data, and when the processor executes the instructions stored in the memory, the method described in the second aspect or any implementation of the second aspect can be implemented.
  • the communication device may also include a memory.
  • the communication device may also include a communication interface, the communication interface is used for the device to communicate with other devices.
  • the communication interface may be a transceiver, a hardware circuit, a bus, a module, a pipe
  • the communication device may be an access network device, or a device, module, or chip set in the access network device, or a device that can be used in conjunction with the access network device.
  • the communication device may be a positioning device, or a device, module, or chip set in the positioning device, or a device that can be used in conjunction with the positioning device.
  • the present application provides a communication system, including a first network element and a second network element.
  • a communication system including a first network element and a second network element.
  • the interaction between the first network element and the second network element is as follows:
  • the second network element sends first information to the first network element, where the first information is used to determine the validity of the candidate training data collected by the first network element, where the determination result of the validity includes valid or invalid;
  • the first network element receives the first information from the second network element
  • the first network element collects candidate training data for the AI model
  • the first network element sends second information to the second network element according to the candidate training data and the first information, where the second information indicates a determination result of validity;
  • the second network element receives the second information from the first network element.
  • the communication system includes a terminal device and an access network device.
  • the communication system includes a terminal device, an access network device, and a positioning device.
  • the terminal device is a location reference device
  • the positioning device is a LMF network element.
  • the present application provides a communication system, comprising a communication device as described in the third aspect or the fifth aspect, and a communication device as described in the fourth aspect or the sixth aspect.
  • the present application further provides a computer program, which, when executed on a computer, enables the computer to execute the method provided in the first aspect, the second aspect, or any implementation of the first aspect or the second aspect.
  • the present application also provides a computer program product, comprising instructions, which, when executed on a computer, enable the computer to execute the method provided in the first aspect, the second aspect, or any implementation of the first aspect or the second aspect.
  • the present application also provides a computer-readable storage medium, in which a computer program or instruction is stored.
  • a computer program or instruction is stored.
  • the computer program or instruction When the computer program or instruction is run on a computer, the computer executes the above-mentioned first aspect, second aspect, or the method provided in any implementation of the first aspect or the second aspect.
  • the present application also provides a chip, which is used to read a computer program stored in a memory to execute the method provided in the above-mentioned first aspect, second aspect, or any aspect of the first aspect or second aspect; or, the chip includes a circuit for executing the above-mentioned first aspect, second aspect, or any aspect of the method provided in the first aspect or second aspect.
  • the present application further provides a chip system, which includes a processor for supporting a device to implement the above-mentioned first aspect, second aspect, or the method provided in any one of the first aspect or second aspect.
  • the chip system also includes a memory, which is used to store the necessary programs and data of the device.
  • the chip system can be composed of a chip, or it can include a chip and other discrete devices.
  • Figure 1 is a schematic diagram of the neural network iteration process.
  • FIG. 2 is a schematic diagram of the architecture of a communication system applicable to an embodiment of the present application.
  • Figure 3 is a schematic flowchart of the method for obtaining training data in AI model training provided in this application.
  • FIG4 is a schematic diagram of a CSI feedback mechanism based on an AI model.
  • FIG5 is an example of obtaining training data from CSI feedback based on the AI model provided in this application.
  • FIG6 is a schematic diagram of the technical solution provided in the present application in an uplink positioning scenario based on an AI model.
  • FIG. 7 is an example of obtaining training data in uplink positioning based on the AI model provided in the present application.
  • FIG8 is a schematic diagram of the technical solution provided in the present application in a downlink positioning scenario based on an AI model.
  • FIG9 is an example of obtaining training data in downlink positioning based on the AI model provided in the present application.
  • FIG10 is a schematic diagram of the AI-assisted sparse beam scanning process.
  • FIG11 is an example of obtaining training data in beam management based on the AI model provided in the present application.
  • FIG12 is a schematic structural diagram of the communication device provided in the present application.
  • FIG13 is a schematic structural diagram of the communication device provided in the present application.
  • AI model refers to a function model that maps an input of a certain dimension to an output of a certain dimension, and its model parameters are obtained through machine learning training.
  • a and b are parameters of the AI model, which can be obtained through machine learning training.
  • the AI models mentioned in the embodiments below in this application are not limited to neural networks, linear regression models, decision tree models, support vector machines (SVM), Bayesian networks, Q learning models or other machine learning (ML) models.
  • Training data set Data used for model training, verification, and testing in machine learning. The quantity and quality of data will affect the effect of machine learning.
  • Training data can include the input of the AI model, or the input and target output of the AI model.
  • the target output is the target value of the output of the AI model, which can also be called the true output value, true output value, label, or label sample.
  • Model training The process of selecting a suitable loss function and using an optimization algorithm to train the model parameters so that the value of the loss function is less than the threshold, or the value of the loss function meets the target requirements.
  • AI model design mainly includes data collection (for example, collecting training data and/or inference data), model training and model inference. It can also include the application of inference results.
  • data collection link the data source is used to provide training data sets and inference data.
  • model training link the AI model is obtained by analyzing or training the training data provided by the data source. Among them, the AI model represents the mapping relationship between the input and output of the model. Learning the AI model through the model training node is equivalent to learning the mapping relationship between the input and output of the model using the training data.
  • the AI model trained through the model training link is used to perform inference based on the inference data provided by the data source to obtain the inference result.
  • This link can also be understood as: inputting the inference data into the AI model, and obtaining the output through the AI model, which is the inference result.
  • the inference result can indicate: the configuration parameters used (executed) by the execution object, and/or the operation performed by the execution object.
  • the reasoning results are published in the reasoning result application link.
  • the reasoning results can be uniformly planned by the execution (actor) entity.
  • the execution entity can send the reasoning results to one or more execution objects (for example, core network equipment, access network equipment, or terminal equipment, etc.) for execution.
  • the execution entity can also feedback the performance of the model to the data source to facilitate the subsequent implementation of the model update training.
  • Loss function It is used to measure the difference or gap between the model's predicted value and the true value.
  • Model application Use the trained model to solve practical problems.
  • Machine learning is an important technical approach to achieve artificial intelligence (AI).
  • Machine learning can be divided into supervised learning, unsupervised learning, and reinforcement learning.
  • supervised learning uses a machine learning algorithm to learn the mapping relationship from sample values to sample labels based on the collected sample values and sample labels, and uses a machine learning model to express the learned mapping relationship.
  • the process of training a machine learning model is the process of learning this mapping relationship.
  • a noisy received signal is a sample
  • the real constellation point corresponding to the signal is a label.
  • Machine learning expects to learn the mapping relationship between samples and labels through training, that is, to make the machine learning model learn a signal detector.
  • the model parameters are optimized by calculating the error between the model's predicted value and the real label.
  • the learned mapping relationship can be used to predict the sample label of each new sample.
  • the mapping relationship learned by supervised learning can include linear mapping and nonlinear mapping. According to the type of label, the learning task can be divided into classification task and regression task.
  • FIG. 1 is a schematic diagram of the neural network iteration process.
  • n samples are selected to form a batch, and then the batch is thrown into the neural network to get the output result. Then the output result and the sample label are thrown into the loss function to calculate the loss of this round. Finally, the derivative of each parameter is combined with the step size parameter to update the parameter.
  • Batch means "batch", which means that the neural network processes data in batches. Batch size is the number of samples processed in each batch. Therefore, generally finding a sample size of appropriate size can speed up the training speed by parallel calculation, and the amount of data processed at one time will not be too large.
  • the training data set is a collection of training samples. Each training sample is an input to the neural network.
  • the training data set is used for model training.
  • the training data set is one of the most important parts of machine learning.
  • the training process of machine learning is essentially to learn certain features from the training data set, so that the output of the neural network is close to the ideal target value (that is, the label or output true value) under the training data set. The difference is minimal.
  • the weights and outputs of the neural networks trained with different training data sets are different. Therefore, the composition and selection of the training data set determine the performance of the trained neural network to some extent.
  • AI models are applied to air interface technology, whether it is offline model update/training or online model update/training, it is necessary to collect data from the real deployment network to form the training data set required for model update/training.
  • a good training data set helps wireless communication AI algorithm design to achieve greater performance gains and improve the generalization ability and robustness of the final design algorithm in various scenarios.
  • AI models are applied to some application scenarios of air interface technology, if the AI model training network element and the training data collection network element are not in the same network element, the AI model training network element and the training data collection network element need to interact with each other for training data. Based on the current technical status, the interaction of training data is usually periodic or continuous, which easily leads to waste of air interface resources.
  • the process of training data collection by the training data collection network element is not constrained by the needs of the AI model training network element, and invalid collection often occurs.
  • the training data collected by the training data collection network element is not the training data actually required by the AI model training network element, resulting in some invalid interactions and waste of air interface resources.
  • the AI model training network element uses these training data for AI model training, it is easy to pollute the training data set of the AI model, resulting in inaccurate gain evaluation, model overfitting, weak generalization ability, poor scene adaptability and other problems.
  • the present application provides a method for obtaining training data in AI model training, which is beneficial to solving or improving the above problems.
  • the communication system can be a fourth generation (4G) communication system (such as a long term evolution (LTE) system), a fifth generation (5G) communication system, a world-wide interoperability for microwave access (WiMAX) or a wireless local area network (WLAN) system, a satellite communication system, or a future communication system, such as a 6G communication system, or a fusion system of multiple systems.
  • 4G fourth generation
  • 5G fifth generation
  • WiMAX world-wide interoperability for microwave access
  • WLAN wireless local area network
  • satellite communication system a satellite communication system
  • a future communication system such as a 6G communication system
  • the 5G communication system can also be called a new radio (NR) system.
  • NR new radio
  • a network element in a communication system may send a signal to another network element, or receive a signal from another network element.
  • the signal may include information, signaling, or data, etc.
  • the network element may also be replaced by an entity, a network entity, a device, a communication device, a communication module, a node, a communication node, etc., and the network element is used as an example for description in this application.
  • the communication system applicable to the present application may include a first network element and a second network element, and optionally, further include a third network element, wherein the number of the first network element, the second network element and the third network element is not limited.
  • FIG. 2 is a schematic diagram of the architecture of a communication system applicable to an embodiment of the present application.
  • FIG. 2 (a) is a schematic diagram of the architecture of a communication system applicable to an embodiment of the present application.
  • the communication system includes a network device 110, a terminal device 120 and a terminal device 130.
  • the terminal devices 120 and 130 can access the network device 110 and communicate with the network device 110.
  • the network device 110 can be an access network device.
  • the communication system may also include an AI entity, and the network device may forward the data related to the AI model reported by the terminal device to the AI entity, and the AI entity performs AI-related operations such as training data set construction and model training, and provides the output of AI-related operations such as the trained AI model, model evaluation, and test results to the network device.
  • the AI entity may also be located inside the network device 110, that is, a module of the network device 110.
  • FIG. 2 (b) is a schematic diagram of the architecture of another communication system applicable to an embodiment of the present application.
  • the communication system includes a network device 110, a terminal device 120, a terminal device 130 and a positioning device 140. Among them, the positioning device 140 and the network device 110 can communicate through interface messages.
  • the positioning device 140 is a location management function (LMF), and the network device 110 can be an access network device, such as a gNB or an eNB, etc., without limitation.
  • the access network device 110 is a gNB
  • the gNB and the LMF can exchange information through NR positioning protocol A (NR positioning protocol A, NRPPa) messages
  • the access network device 110 is an eNB
  • the eNB and the LMF can exchange information through LTE positioning protocol (LTE positioning protocol, LPP) messages.
  • the terminal device and the positioning device 140 can also communicate directly, such as the interaction between the terminal device 130 and the positioning device 140 shown in (b) of Figure 2.
  • the AI entity can be configured inside the positioning device 140, or separately configured from the positioning device 140, without limitation.
  • the positioning device and the network device can be different modules of the same device, or they can be separate different devices.
  • a network device can serve one or more terminal devices at the same time.
  • a terminal device can also access one or more network devices at the same time.
  • the embodiment of the present application does not limit the number of terminal devices and network devices included in the wireless communication system.
  • the positioning device 140 of (b) of Figure 2 is not limited to being an LMF network element, but can also be other network elements with positioning functions, and the number thereof is not limited.
  • a network device may be a device with wireless transceiver functions, and the network device may be a device that provides wireless communication function services, and is usually located on the network side, including but not limited to the next generation base station (gNodeB, gNB) in the fifth generation (5th generation, 5G) communication system, the base station in the sixth generation (6th generation, 6G) mobile communication system, the base station in the future mobile communication system, or the access node (access point, AP) in the wireless fidelity (wireless fidelity, WiFi) system, the evolved node B (evolved node B, eNB) in the long term evolution (long term evolution, LTE) system, the wireless
  • the network device may include a radio network controller (RNC), a node B (NB), a base station controller (BSC), a home base station (e.g., home evolved NodeB, or home Node B, HNB), a base band unit (BBU), a transmission reception point (TRP), a transmitting point (TP), a base
  • the network device may include a centralized unit (CU) node, or a distributed unit (DU) node, or a RAN device including a CU node and a DU node, or a RAN device including a control plane CU node, a user plane CU node, and a DU node, or the network device may also be a wireless controller, a relay station, a vehicle-mounted device, and a wearable device in a cloud radio access network (CRAN) scenario.
  • the base station can be a macro base station, a micro base station, a relay node, a donor node, or a combination thereof.
  • the base station can also refer to a communication module, a modem, or a chip used to be set in the aforementioned device or apparatus.
  • the base station can also be a mobile switching center and a device that performs the base station function in device-to-device (D2D), vehicle-to-everything (V2X), and machine-to-machine (M2M) communications, a network-side device in a 6G network, and a device that performs the base station function in future communication systems.
  • the base station can support networks with the same or different access technologies without limitation.
  • the network equipment may be fixed or mobile.
  • the access network equipment 110 may be stationary and responsible for wireless transmission and reception in one or more cells from the terminal devices 120 and 130.
  • the access network equipment 110 may also be mobile, for example, a helicopter or a drone may be configured to act as a mobile base station, and one or more cells may move according to the location of the mobile base station. It should be understood that in other examples, a helicopter or a drone may be configured to be used as a device for communicating with the base station 110.
  • the communication device used to implement the above access network function can be an access network device, or a network device with some functions of accessing the network, or a device capable of supporting the implementation of the access network function, such as a chip system, a hardware circuit, a software module, or a hardware circuit plus a software module, which can be installed in the access network device or used in combination with the access network device.
  • the communication device used to implement the access network device function is an access network device for example.
  • the terminal device can be an entity on the user side for receiving or transmitting signals, such as a mobile phone.
  • the terminal device includes a handheld device with a wireless connection function, other processing devices connected to a wireless modem, or a vehicle-mounted device.
  • the terminal device can be a portable, pocket-sized, handheld, computer-built-in, or vehicle-mounted mobile device.
  • the terminal device 120 can be widely used in various scenarios, such as cellular communication, WiFi system, D2D, V2X, peer to peer (P2P), M2M, machine type communication (MTC), Internet of Things (IoT), virtual reality (VR), augmented reality (AR), industrial control, automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, drone, robot, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, etc.
  • cellular communication WiFi system
  • D2D peer to peer
  • M2M machine type communication
  • IoT Internet of Things
  • VR virtual reality
  • AR augmented reality
  • industrial control automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, drone, robot, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, etc.
  • Some examples of the communication device 120 include: user equipment (UE) of the 3GPP standard, a station (STA) in a WiFi system, a fixed device, a mobile device, a handheld device, a wearable device, a cellular phone, a smart phone, a session initialization protocol (SIP) phone, a laptop, a personal computer, a smart book, a vehicle, a satellite, a global positioning system (GPS) device, a target tracking device, a drone, a helicopter, an aircraft, a ship, a remote control device, a smart home device, an industrial device, a personal communication service (PCS) phone, a wireless local loop (WLL) station, a personal digital assistant (PDA), a wireless network camera, a tablet computer, a handheld computer, a mobile internet device (MID), a wearable device such as a smart watch, a virtual reality (VR) device, an augmented reality (AR) device, a wireless terminal in industrial control, a terminal in a vehicle networking system, a wireless
  • the terminal device 120 may be a wireless device in the above scenarios or a device used to be set in a wireless device, such as a communication module, a modem or a chip in the above device.
  • the terminal device may also be referred to as a terminal, user equipment (UE), a mobile station (MS), a mobile terminal (MT), etc.
  • the terminal device may also be a terminal device in a future wireless communication system.
  • the terminal device may also include a location reference device, such as an automatic navigation device.
  • the embodiment of the present application does not limit the specific technology and specific device form adopted by the terminal device.
  • the communication device used to implement the functions of the terminal device can be a terminal device, or a terminal device having some functions of the above communication device, or a device capable of supporting the functions of the above terminal device, such as a chip system, which can be installed in the terminal device or used in combination with the terminal device.
  • the chip system can be composed of a chip, or it can include a chip and other discrete devices.
  • the number and type of each device in the communication system shown in Figure 2 are for illustration only, and the present application is not limited to this.
  • the communication system may also include more terminal devices, more access network devices, more positioning devices, and other network elements, such as core network devices, and/or network elements for implementing artificial intelligence functions.
  • the first network element may be a network element that collects training data of the AI model
  • the second network element is a training network element of the AI model
  • the second network element may be a training network element of the AI model and also a network element where AI reasoning occurs.
  • the first network element and the second network element may be logically deployed separately.
  • the first network element and the second network element may be physically deployed in the same network element or different network elements, without limitation.
  • the first network element receives first information from the second network element, where the first information indicates a determination of validity of candidate training data collected by the first network element, wherein the determination result of the validity includes valid or invalid.
  • the first network element may determine the validity of the collected candidate training data based on the first information. In other words, based on the first information, the first network element may determine whether the collected candidate training data is valid.
  • the first network element determines that the collected candidate training data contains valid training data based on the first information
  • the first network element determines that the collected candidate training data is valid; if the first network element determines that the collected candidate training data does not contain valid candidate training data based on the first information, the first network element determines that the collected candidate training data is invalid.
  • the determination result is valid.
  • the valid candidate training data becomes the training data collected by the first network element this time.
  • the valid candidate training data (hereinafter referred to as "valid data”) may be part or all of the collected candidate training data, without limitation, and are collectively referred to as first training data in this article. If the candidate training data collected by the first network element does not contain valid candidate training data, the determination result is invalid.
  • the first information indicates a constraint condition, which is used by the first network element to determine the validity of the collected candidate training data.
  • constraints include one or more of the following:
  • the first information is used to determine the constraint condition.
  • the first information indicates one or more of the following information:
  • the constraint condition includes one or more quality indicators
  • the first information indicates a threshold of the one or more quality indicators, and the determination criteria of the one or more quality indicators are predefined by a protocol.
  • the first network element determines the constraint condition according to the first information and the protocol predefined.
  • the first information indicates a threshold of the one or more quality indicators and a determination criterion of the one or more quality indicators.
  • the first network element determines the constraint condition according to the first information.
  • the constraint condition includes multiple quality indicators
  • the first information indicates that some of the quality indicators
  • the threshold of the quantity indicator, the threshold of another part of the quality indicators and the judgment criteria of the multiple quality indicators are predefined by the protocol.
  • the first network element determines the constraint condition according to the first information and the protocol predefined.
  • the first information indicates a threshold of some quality indicators and an index information
  • the index information is used to determine the judgment criteria of the some quality indicators, as well as the thresholds of other quality indicators in the constraint conditions and the judgment criteria of the other quality indicators.
  • the first network element determines the constraint conditions based on the first information and the index information.
  • the first information indicates an index information
  • the index information is used to determine a threshold of one or more quality indicators and a determination criterion of the one or more quality indicators.
  • the first network element determines the constraint condition according to the index information.
  • the "protocol pre-definition" in the above example may also be other implementation methods such as pre-configuration or pre-storage, without limitation.
  • the first information further indicates a maximum number k of validity determinations, where k is a positive integer.
  • the first information includes multiple items of the above information
  • the multiple items of information can be carried in one message or carried in multiple messages respectively, that is, the first information can be carried in one message or carried in multiple messages.
  • the maximum duration of collecting candidate training data corresponding to a single validity determination is denoted as Z below, where Z is a number greater than 0.
  • the maximum duration of collecting the candidate training data that is, the maximum duration that the candidate training data can be used for validity determination. If the retention duration of the candidate training data exceeds the maximum duration Z, the candidate training data becomes invalid and is no longer used for validity determination.
  • the maximum duration Z may be the same as the interval between two adjacent validity determinations, or may be greater than or less than the interval between two adjacent validity determinations.
  • the candidate training data corresponding to one validity determination may include all or part of the candidate training data corresponding to one or more validity determinations before the validity determination.
  • the interval between the two adjacent validity determinations can be fixed, that is, the validity determination is performed periodically within a certain time, or it can be variable, that is, the validity determination time is not fixed.
  • the validity determination is to count the candidate training data that meet the threshold requirements.
  • the validity determination is completed and the determination result is valid; when the number of candidate training data that meet the threshold requirements does not meet the requirements and exceeds the maximum interval time T of the validity determination (that is, the preset interval time threshold) or the number of candidate training data collected exceeds the preset threshold (that is, the maximum number of candidate training data collected), the validity determination is also completed and the determination result is invalid.
  • Specific time-related information of the validity determination such as the determination time, the start time of the periodic determination, or one or more of the period, the maximum interval time T, the maximum number of candidate training data collected, etc., can be fully or partially predefined by the protocol, or based on the configuration.
  • the second network element sends the first information to the first network element, and the first information is used for the validity determination of the candidate training data collected by the first network element.
  • the second network element indicates the requirements of the training data to the second network element through the first information.
  • only the candidate training data that meets the requirements can be used as training data for the training or update of the AI model. It can be seen that the candidate training data collected by the first network element needs to be "screened" before the candidate training data that meets the requirements can be used as training data and provided by the first network element to the second network element for use. Therefore, after the first network element collects the candidate training data, it will determine the validity of the collected candidate training data according to the first information.
  • the determination result is invalid, it means that the candidate training data collected this time is not required by the second network element, that is, the candidate training data collected this time does not contain candidate training data that meets the requirements.
  • the first network element may involve the re-collection of training data. Therefore, in the process of the first network element collecting training data for the AI model, the training data may not be obtained by collecting it once.
  • the maximum time interval T for a single validity determination is equivalent to specifying how often the first network element performs a validity determination.
  • the first network element collects the candidate training data again. After a period of time, the first network element performs the i+1th validity judgment on the collected candidate training data, where i is a positive integer. Therefore, in an embodiment of the present application, the number of times the first network element collects candidate training data for the AI model and the number of times the first network element performs a validity judgment are corresponding, or equal. In other words, each time the first network element performs a validity judgment, it represents a collection of candidate training data before this judgment. For the sake of clarity in the description of the technical solution, the collection of candidate training data before the i-th judgment is referred to as the i-th collection.
  • the i+1th validity judgment can be for the candidate training data obtained in the i+1th collection, or it can be for the candidate training data obtained in the i+1th collection and one or more collections before the i+1th collection, without limitation. In this implementation, it may involve how the first network element handles the problem after a validity judgment. This is a question about the candidate training data collected before the judgment.
  • the first network element After the first network element starts the i-th collection of candidate training data, after a time interval T0 (less than or equal to the maximum time interval T), the first network element performs a validity determination on the candidate training data collected for the i-th time, that is, the i-th validity determination. Assuming that the determination result of the i-th validity determination is invalid, the first network element can discard the candidate training data collected for the i-th time, and perform the i+1-th collection again if the maximum number of validity determinations k is not exceeded. In this example, each validity determination is only for the candidate training data collected within the time interval T0, and when the collection is invalid, the candidate training data collected this time is discarded.
  • a validity determination can be for candidate training data collected within multiple time intervals T0.
  • the candidate training data for a validity determination may come from multiple collections.
  • the first network element can retain part of the candidate training data collected for the i-th time. For example, the first network element retains the part of the candidate training data collected for the i-th time that meets the determination criteria of some quality indicators in the constraint conditions. After that, the i+1th collection is performed.
  • the first network element After a time interval T0, the first network element performs a validity judgment on the candidate training data collected for the i+1th time and the part of the candidate training data collected for the ith time whose retention time does not exceed the maximum time Z, that is, the i+1th validity judgment.
  • the first network element can retain the candidate training data that meets the judgment criteria of some quality indicators in the candidate training data collected each time, and after completing a new collection, the retained historical candidate training data that meets the judgment criteria of some quality indicators and the newly collected candidate training data are judged for validity together. It can be seen that in a validity judgment, the candidate training data determined to be invalid is for that validity judgment, and does not mean that the candidate training data determined to be invalid in this validity judgment can never be used as training data.
  • the present application does not limit these specific implementation methods.
  • the constraint condition is based on the application scenario of the AI model.
  • the application scenario of the AI model includes but is not limited to the following scenarios:
  • one or more of the quality indicators involved in the constraints, the threshold of the quality indicators, the judgment criteria of the quality indicators, the threshold of the number of training data that meets the judgment criteria of the quality indicators, and the judgment criteria of the number of training data may be different. Examples will be given below for different application scenarios.
  • the first network element collects candidate training data for the AI model.
  • the first network element collects candidate training data for the AI model.
  • the first network element may start collecting candidate training data for the AI model after receiving the first information, that is, based on the triggering of the first information.
  • the first network element may also start collecting candidate training data for the AI model before or at the same time as receiving the first information. That is, the order of occurrence of step 310 and step 320 may not be limited.
  • the second network element may send an updated first information to the first network element.
  • the update of the first information mainly refers to the update of the constraint conditions determined by the first information.
  • the validity of the collected candidate training data is determined based on the constraint conditions determined by the updated first information.
  • only the first information received by the first network element at a certain time is used as an example to illustrate the validity determination and subsequent processes.
  • the first network element sends second information to the second network element, where the second information indicates a determination result of validity of the candidate training data collected by the first network element.
  • the first network element determines that the collected candidate training data is valid according to the first information
  • the first network element sends second information to the second network element, and the second information indicates that the candidate training data collected by the first network element is valid.
  • the first network element sends the first training data to the second network element, and the first training data itself implicitly indicates that the candidate training data collected by the first network element is valid.
  • the first network element sends the first training data to the second network element, and at this time, the first network element also sends information indicating that the candidate training data collected by the first network element is valid, for example, information a.
  • the first network element sends the first training data and information a to the second network element, and information a indicates that this collection is valid.
  • the first network element determines that the collected candidate training data is invalid according to the first information
  • the first network element sends second information to the second network element, and the second information indicates that the candidate training data collected by the first network element is invalid. It should be understood that in the case where the candidate training data collected by the first network element is invalid, the first network element only sends an indication that the collected candidate training data is invalid to the second network element, and does not send the collected invalid candidate training data, thereby reducing the waste of air interface resources.
  • the first network element determines that the collected candidate training data is invalid, the first network element discards the candidate training data collected this time; or, in some implementations described above, in a validity determination, the invalid candidate training data may also be discarded. The data is retained for subsequent validity determination. Further, if the second network element instructs the first network element to re-collect the training data of the AI model, the first network element re-collects the candidate training data of the AI model.
  • the first network element collects candidate training data for the AI model, specifically, the first network element measures a reference signal from a second network element or a third network element to obtain candidate training data for the AI model, or in other words, the candidate training data includes a measurement result obtained by measuring the reference signal by the first network element.
  • a reference signal generally refers to a signal used for channel measurement.
  • the channel measurement can be used for one or more of the functions of channel state information feedback, beam management, or positioning.
  • the reference signal may include a channel state information reference signal, a synchronization signal, such as a primary synchronization signal and/or a secondary synchronization signal, a physical broadcast signal, a synchronization signal and a physical broadcast signal block (SSB), a demodulation reference signal, a phase tracking reference signal, or one or more of a positioning reference signal.
  • a synchronization signal such as a primary synchronization signal and/or a secondary synchronization signal
  • a physical broadcast signal such as a primary synchronization signal and/or a secondary synchronization signal
  • SSB physical broadcast signal block
  • demodulation reference signal such as a primary synchronization signal and/or a secondary synchronization signal
  • SSB physical broadcast signal block
  • the reference signal may be different, and the following embodiments will be illustrated for different application scenarios.
  • the third network element refers to a network element different from the second network element.
  • the first network element measures the reference signal from the second network element and obtains the measurement result.
  • the measurement result may be one or more.
  • the candidate training data of the AI model collected by the first network element includes the one or more measurement results.
  • the second network element before the second network element sends the reference signal to the first network element, the second network element sends air interface transmission configuration information to the first network element, and the air interface transmission configuration information corresponds to the air interface transmission configuration, and the air interface transmission configuration information indicates that the first network element collects candidate training data for the AI model based on the air interface transmission configuration.
  • the second network element sends a reference signal according to the air interface transmission configuration, and the first network element measures the reference signal from the second network element to obtain the measurement result, thereby collecting candidate training data based on the air interface transmission configuration.
  • the first network element is a UE
  • the second network element is an access network device, such as a base station.
  • the first network element measures a signal from a third network element and obtains a measurement result.
  • the measurement result may be one or more.
  • the candidate training data of the AI model collected by the first network element includes the one or more measurement results.
  • the first network element is an access network device, such as a base station, and the third network element is a UE.
  • the first network element configures the third network element to send a reference signal.
  • the first network element sends air interface transmission configuration information to the third network element, and the air interface transmission configuration information corresponds to the air interface transmission configuration.
  • the third network element sends a reference signal based on the air interface transmission configuration, and the first network element measures the reference signal from the third network element to obtain a measurement result, thereby obtaining candidate training data based on the air interface transmission configuration.
  • the air interface transmission configuration may include one or more of the following:
  • the number of antenna ports used by the reference signal is the number of antenna ports used by the reference signal
  • the frequency domain density of the reference signal or,
  • the first network element collects candidate training data for the AI model and determines the validity of the candidate training data based on the first information, which can be understood as screening the collected candidate training data. If the collected candidate training data is valid, the first network element sends the valid candidate training data to the second network element for the second network element to train or update the AI model. At this time, the valid candidate training data is the training data. If the candidate training data collected by the first network element is invalid, the first network element indicates to the second network element that the collection of the candidate training data is invalid.
  • the candidate training data collected once is invalid, it also means that no training data is collected this time. If the candidate training data collected once is valid, it also means that the training data is collected this time. At this time, the valid candidate training data becomes the training data, referred to as the first training data in this article, and is provided by the first network element to the second network element.
  • the first network element re-collects the training data of the AI model also means “the first network element re-collects the candidate training data of the AI model”. This is because if the first network element fails to collect the training data, it will try to re-collect it, and the purpose of re-collection is to collect the training data of the AI model, but the process of collecting the training data of the AI model is to first collect the candidate training data and then filter the training data from the candidate training data.
  • the second network element After receiving the second information from the first network element, if the second network element determines based on the second information that the current collection by the first network element is invalid, in one possible case, the second network element determines that the training data of the AI model needs to be collected again.
  • the second network element sends the third information to the first network element, and the third information instructs the first network element to re-collect the training data of the AI model.
  • the third information indicates the maximum number of validity determinations k.
  • each validity determination includes a new batch of training data, that is, a new collection of training data sets, and thus the maximum number of validity determinations can also be referred to as execution times. The maximum number of times to collect training data sets.
  • the second network element when the second network element receives an instruction from the first network element that the candidate training data collected by the first network element is invalid, the second network element sends a third message to the first network element to instruct the first network element to re-collect the training data of the AI model.
  • the third message indicates the maximum number of validity determinations k, where k is a positive integer.
  • the first network element re-collects or continues to collect the candidate training data according to the third information. Re-collection or continued collection may involve multiple times, and a validity determination is performed after each re-collection or continued collection is completed.
  • the first network element can continue with the next re-collection and the next validity determination until the maximum number of validity determinations k is reached. If the determination results of the 1st validity determination to the k-1th validity determination are all invalid, and the determination result of the kth validity determination is still invalid, the first network element will stop collecting training data.
  • the second network element when the second network element instructs the first network element to re-collect training data through the third information, the second network element may also indicate the maximum number of validity determinations k to the first network element. That is, the maximum number of validity determinations k is sent after the second network element determines that the training data needs to be re-collected.
  • the maximum number of validity determinations k may be included in the third information, or carried by other information other than the third information.
  • the second network element indicates the maximum number of validity determinations k in the first information sent to the first network element.
  • the second network element constrains the process of the first network element re-collecting candidate training data, so that the first network element will not fall into an unlimited time-limited re-collection cycle when the training data (that is, valid candidate training data) is not collected, but stops collecting after the maximum number of validity determinations k is reached, regardless of whether the training data is collected.
  • the first network element Before exceeding the maximum number of validity judgments k, if the first network element determines, based on the first information, that the result of the j-th validity judgment is valid, that is, the candidate training data targeted by the j-th validity judgment includes valid candidate training data, the first network element sends fourth information to the second network element, the fourth information includes the second training data, and the fourth information indicates that the result of the j-th validity judgment is valid, and the second training data specifically may include the valid candidate training data in the candidate training data targeted by the j-th validity judgment, j is less than or equal to k, and j is a positive integer.
  • the j-th validity determination can be considered as a validity determination for a set of candidate training data, and all candidate training data included in the set are the training data for the j-th validity determination.
  • the training data for the j-th validity determination is not limited to the candidate training data collected for the j-th time, but may also include candidate training data obtained from one or more collections before the j-th collection, without limitation.
  • the first network element sends the second training data and information a to the second network element, where the information a indicates that this collection is valid.
  • the maximum number of validity determinations k corresponds to a starting time
  • the starting time should be understood as the starting time of the collection process of the training data corresponding to the maximum number of validity determinations k.
  • the starting time may be the time when the first network element receives the first information or the third information. Equivalently, the first network element starts collecting training data for the AI model from the moment the first information or the third information is received.
  • the end time of the collection process is uncertain.
  • the collection process ends, and the first network element sends valid candidate training data (that is, the first training data) to the second network element, j is less than or equal to k, and j is a positive integer.
  • the determination results from the first validity determination to the kth validity determination are all invalid, the moment when the determination result of the kth validity determination is determined to be invalid is the end time of the collection process.
  • the method for obtaining training data provided in the present application is that the training data collecting network element determines the validity of the collected candidate training data and provides valid candidate training data to the training network element of the AI model, thereby ensuring that the collecting network element only provides the training network element with training data that meets the requirements. Since the training data that does not meet the requirements is filtered out on the collecting network element side, the interaction of invalid training data is eliminated, which not only saves air interface resources, but also avoids the pollution of the training data set on the training network element side and avoids other adverse effects caused thereby.
  • the above describes in detail the main process of the method for obtaining training data in the AI model.
  • the following is an example of the method for obtaining training data when the AI model is applied in different scenarios.
  • CSI Channel state information
  • the training or updating of the AI model in application scenario 1 is deployed on the access network device side.
  • the access network device sends a downlink reference signal to the UE so that the UE can obtain a measurement result by measuring the downlink reference signal.
  • the measurement result is the candidate training data.
  • the obtained candidate training data is judged for validity, and the valid candidate training data is provided to the access network device side for training or updating the AI model.
  • the downlink reference signal can be specifically CSI-RS.
  • the valid candidate training data provided by the UE to the access network device is the label of the AI model, specifically CSI.
  • the access network side needs to obtain downlink CSI to determine one or more of the configurations such as resources, modulation and coding scheme (MCS) and precoding of the downlink data channel for scheduling the UE.
  • TDD time division duplex
  • the access network device can obtain the uplink CSI by measuring the uplink reference signal, and then infer the downlink CSI, for example, using the uplink CSI as the downlink CSI.
  • FDD frequency division duplex
  • the downlink CSI is obtained by the UE measuring the downlink reference signal.
  • the UE measures the CSI-RS or the synchronizing signal and physical broadcast channel block (SSB) and other signals to obtain the downlink CSI.
  • the UE generates a CSI report in a manner predefined by the protocol or preconfigured by the access network device, and feeds the downlink CSI back to the access network device through the CSI report, so that the access network device obtains the downlink CSI.
  • SSB physical broadcast channel block
  • FIG4 is a schematic diagram of a CSI feedback mechanism based on an AI model.
  • the auto encoder (AE) model is composed of two sub-models, an encoder and a decoder.
  • AE generally refers to a network structure composed of two sub-models.
  • the AE model can also be called a bilateral model, a dual-end model or a collaborative model.
  • the encoder and decoder of AE are usually trained together and can be used in conjunction with each other.
  • CSI feedback can be implemented based on the AI model of AE. For example, the UE side measures the downlink reference signal sent by the base station to obtain the measured CSI.
  • the UE compresses and quantizes the measured CSI by the encoder, and feeds back the compressed and quantized information to the base station, such as the “feedback CSI information” shown in FIG3 .
  • the base station recovers the “feedback CSI information” through the decoder to obtain the recovered CSI.
  • the input of the decoder is the CSI information fed back by the UE, and the training of the decoder requires the CSI obtained by the UE measurement as the true value (or label) of the recovered CSI.
  • the AI model deployed on the access network device side can be a decoder as shown in Figure 4.
  • the access network device determines that it is necessary to collect training data for the AI model.
  • the access network device sends first information to the UE, where the first information is used to determine the validity of the candidate training data collected by the UE.
  • the determination result of the validity may be valid or invalid.
  • the first information indicates a constraint condition for determining validity of candidate training data collected by the UE.
  • step 310 For the first information and constraint conditions, etc., please refer to the relevant description in step 310, which will not be repeated here.
  • the quality indicator of the measurement result may be the SINR of the training data and the number of training data.
  • the first information indicates the threshold Q of the SINR and the threshold N of the number of training data, and the SINR determination criterion and the number of training data determination criterion (for example, the SINR is greater than or equal to Q, and the number of training data is greater than or equal to N) may be predefined by the protocol.
  • the first information indicates the threshold Q of the SINR and the threshold N of the number of training data, as well as the SINR determination criterion and the number of training data determination criterion.
  • the first information indicates the threshold Q and the threshold N, and the first information includes an information field for indicating the determination criterion.
  • the information field includes 1 bit, and the 1 bit corresponds to the SINR determination criterion and the number of training data determination criterion, for example, the value of the 1 bit is "1" indicating that "the SINR of the training data is greater than or equal to Q, and the number of training data is greater than or equal to N", and the value of the 1 bit is "0" indicating that "the SINR of the training data is greater than Q and the number of training data is greater than N”.
  • the information domain includes 2 bits b1 b0 , where b1 corresponds to the SINR judgment criterion, and b0 corresponds to the judgment criterion of the number of training data.
  • the first information indicates the threshold Q of the SINR of the training data and the threshold N of the number of training data.
  • the judgment criterion of the quality indicator of the first information part and the judgment criterion of the other part of the quality indicator are predefined by the protocol.
  • the first information indicates the threshold Q and the threshold N.
  • the first information also includes a 1-bit information domain.
  • the value of the 1 bit When the value of the 1 bit is 1, it means “SINR is greater than or equal to Q", and when the value of the 1 bit is 0, it means “SINR is less than Q"; wherein, the judgment criterion of the number of training data is predefined by the protocol, for example, "the number of training data is at least N". It should be understood that the above implementation is only an example of the first information being used to determine the constraint condition, and is not limited.
  • N can be an integer multiple of a batch during AI model training or the number of training data required for the AI model to converge.
  • the access network device sends a reference signal to the UE.
  • the UE obtains one or more measurement results by measuring the reference signal of the access network device.
  • the measurement result can also be replaced by the measurement result of the reference signal or the channel measurement result. This replacement expression is also applicable to the implementation in other application scenarios. In the examples, no repeated description is given below.
  • the measurement results include a channel response, such as a channel response matrix.
  • the UE may obtain one measurement result through one measurement, and in this case, the candidate training data includes the one measurement result; optionally, the UE may obtain multiple measurement results through multiple measurements, and in this case, the candidate training data includes the multiple measurement results.
  • the quality indicators of the measurement results may include, but are not limited to, one or more of the following: first path power, first path arrival delay, timing error group (TEG), average power of time domain sampling points, phase difference between antenna ports, equivalent SINR of the full band or sub-band, interference level of the full band or sub-band, line of light (LOS) probability, inter-station synchronization error, or, one or more of the confidence level of the measurement results, etc., without limitation.
  • the quality of this indicator is applicable to application scenario 1 or other application scenarios described below, without limitation. It should be understood that these quality indicators can be obtained by performing corresponding processing on the measurement results of the reference signal.
  • the specific processing process is not limited here, for example, it can be some known or future processing.
  • the air interface transmission configuration information indicates the relevant air interface configuration for the access network device to send the reference signal.
  • the air interface transmission configuration information may include but is not limited to one or more of the information such as the transmission power of the reference signal, the number of antenna ports used by the access network device to send the reference signal, the frequency bandwidth of the reference signal, the frequency domain density of the reference signal, and the period of the reference signal.
  • the air interface transmission configuration information may also include other relevant information, which are not listed here one by one.
  • the candidate training data collected by the UE is one or more measurement results obtained by measuring a reference signal from an access network device, or one or more channel measurement results.
  • the reference signal can be a channel state information-reference signal (CSI-RS).
  • CSI-RS channel state information-reference signal
  • the UE determines the validity of the collected candidate training data according to the first information.
  • the candidate training data is a plurality of measurement results obtained by the UE by measuring a reference signal, and the plurality of measurement results are the candidate training data.
  • the UE determines whether the plurality of measurement results contain valid candidate training data (or valid measurement results) according to the constraint condition.
  • the constraint condition is "the quality index (such as SINR) is greater than or equal to the threshold Q, and the number of candidate training data meeting the quality index greater than or equal to the threshold Q is at least N”
  • the UE determines whether the plurality of measurement results collected contain measurement results with a quality equal to or greater than the threshold Q.
  • the measurement result with a quality index equal to or greater than the threshold Q is recorded as measurement result 1 below.
  • the UE determines that the plurality of measurement results collected contain measurement result 1, it is also necessary to determine whether the number of measurement results 1 reaches N. If it is determined that a valid measurement result is collected according to the constraint condition, the UE determines that the candidate training data collected this time is valid, wherein the valid candidate training data (i.e., the first training data, sometimes also referred to as valid data below) is the part of the measurement results that meets the constraint condition. For example, if the number of measurement results 1 is P, where P is an integer greater than or equal to N, then the P measurement results 1 are the valid data collected this time, that is, the first training data.
  • the UE determines that the collected multiple measurement results do not include a measurement result that satisfies the constraint condition, for example, the collected multiple measurement results include measurement result 1 whose SINR is equal to or greater than the threshold Q, but the number of measurement results 1 is less than N; or, the SINRs of the collected multiple measurement results are all less than the threshold Q, in this case, the UE determines that the candidate training data collected this time is invalid.
  • the UE sends second information to the access network device based on the determination of the validity of the collected candidate training data, where the second information indicates the determination result of the validity.
  • the second information indicates that the candidate training data collected by the UE is valid.
  • the second information may be the collected valid candidate training data itself, such as the P measurement results 1 in the above example.
  • the P measurement results 1 are both valid candidate training data and the P measurement results 1 also implicitly indicate that the candidate training data collected by the UE is valid.
  • the UE sends the second information and valid candidate training data.
  • the second information indicates that the candidate training data collected by the UE is valid.
  • the second information may include 1 bit. When the value of the 1 bit is "1", it indicates that the candidate training data collected by the UE is valid.
  • the UE sends valid candidate training data to the access network device.
  • the previous example can further save signaling overhead while being able to indicate that the candidate training data collected by the UE is valid.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the second information may include 1 bit, and when the value of the 1 bit is "0", it indicates that the candidate training data collected by the UE is invalid.
  • the second information may be carried by uplink control information (UCI) signaling, for example, UCI includes 1 bit of information, and the 1 bit is used to indicate whether the candidate training data collected by the UE is valid or invalid.
  • UCI uplink control information
  • the valid candidate training data is Training data can also be sent in UCI without limitation.
  • the access network device determines whether the collection of candidate training data of the UE is valid according to the second information.
  • the second information indicates that the candidate training data collected by the UE is valid, in which case the access network device also obtains the valid candidate training data collected by the UE from the UE. Further, the access network device trains or updates the AI model according to the valid candidate training data, as in step 507.
  • the access network device trains the AI model to obtain the AI model or updates the AI model.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the access network device maintains the CSI feedback of the original AI model, or switches to the CSI feedback of the non-AI model.
  • maintaining the original AI model can be for the scenario where a trained AI model has been deployed on the access network device, and the collection of training data this time is based on the purpose of updating the AI model; switching to a non-AI model can be for the scenario where a trained AI model has not been deployed on the access network device, and the collection of training data this time is for the purpose of training the AI model.
  • the access network device can switch to CSI feedback in non-AI mode.
  • the access network device performs CSI feedback based on the original AI model or switches to a non-AI model.
  • step 507 or step 508 a training data collection process ends.
  • the second information indicates that the candidate training data collected by the UE is invalid, and after the access network device obtains the second information, it determines to collect the training data again, such as steps 509-510.
  • the access network device determines to re-collect the training data of the AI model.
  • the access network device sends third information to the UE, and the third information instructs the UE to re-collect training data of the AI model.
  • the third information further indicates a maximum number of validity determinations k, where k is a positive integer.
  • the maximum number of validity determinations k may also be indicated by the first information, without limitation.
  • the access network device may update the air interface transmission configuration. Accordingly, the UE re-collects the candidate training data of the AI model based on the updated air interface transmission configuration.
  • the access network device sends air interface transmission configuration information to the UE, where the air interface transmission configuration information indicates an updated air interface transmission configuration.
  • the air interface transmission configuration information in step 511 indicates the updated air interface transmission configuration.
  • the update of the air interface transmission configuration may include the update of the transmit power of the reference signal, the number of antenna ports used when the access network device sends the reference signal, the bandwidth of the reference signal, the frequency domain density of the reference signal, the period of the reference signal, etc., without limitation.
  • the update of the air interface transmission configuration includes an increase in the transmit power of the reference signal and an increase in the frequency domain density of the reference signal
  • the access network device sends a reference signal to the UE with a greater transmit power and a greater frequency domain density in an attempt to allow the UE to obtain candidate training data that meets the constraints.
  • the UE when re-collecting the training data of the AI model, the original air interface transmission configuration is not updated.
  • the UE re-collects the candidate training data under the original air interface transmission configuration, and determines the validity of the re-collected candidate training data based on the constraints, and indicates the validity determination result to the access network device.
  • the UE re-collects training data for the AI model.
  • the determination of the validity of the re-collected candidate training data and the indication of the determination result are similar to the above process in Figure 5, and will not be repeated. It should be understood that in the process of re-collecting training data, the UE is constrained by the maximum number of validity determinations k.
  • the maximum number of validity determinations k can be determined by the access network device according to the urgency of training data collection.
  • the urgency can refer to the time since the last update of the AI model. For example, if the time interval since the last update of the AI model is large and exceeds a certain threshold, it is considered that the update demand of the AI model is relatively urgent, because the larger the time interval, the greater the possibility of changes in the channel environment, which means that the matching degree of the AI model to the current channel environment may be reduced, and therefore the update demand is more urgent.
  • the maximum number of validity determinations k can be set larger accordingly, so that after an invalid collection, it is expected to obtain valid candidate training data through multiple re-collections.
  • the urgency determination criteria can also be implemented in other ways, without limitation.
  • FIG 6 is a schematic diagram of the technical solution provided by the present application in an uplink positioning scenario based on an AI model.
  • the training or update deployment of the AI model is performed on the network side.
  • the AI model can be deployed in the positioning device of the core network, such as an LMF network element.
  • the input of the AI model is one or more channel responses (or channel measurement results) corresponding to one or more detection reference signals
  • the output of the AI model is the position of the UE.
  • the transmitting end of the one or more detection reference signals, such as the UE can be one or more
  • the receiving end such as the access network device, can also be one or more.
  • the positioning device trains the AI model used for positioning, it obtains from the access network side a plurality of measurement results obtained by measuring a plurality of detection reference signals by an access network device or by measuring one or more detection reference signals by each of a plurality of access network devices, as well as the location information of a third network element.
  • the aforementioned plurality of detection reference signals may include a plurality of detection reference signals from a third network element, or include one or more detection reference signals from each of a plurality of third network elements.
  • the location information of the third network element at different times when sending a plurality of detection reference signals, or the location information of the plurality of third network elements when each sending one or more detection reference signals at one or more times, is used as the true value (i.e., label) of the location information output by the AI model.
  • the positioning device determines that it is necessary to collect training data for the AI model.
  • the positioning device sends first information to the access network, where the first information is used to determine the validity of candidate training data collected by the access network device.
  • the determination result may be valid or invalid.
  • the first information can be carried by an interface message between the positioning device and the access network device.
  • the positioning device is LMF and the access network device is gNB
  • the first information between LMF and gNB can be included in the NRPPa message.
  • the access network device sends air interface transmission configuration information (such as air interface transmission configuration information #1) to the third network element, where the air interface transmission configuration information indicates the air interface transmission configuration when the third network element sends a sounding reference signal.
  • air interface transmission configuration information such as air interface transmission configuration information #1
  • the third network element is a network element that can provide its own location information.
  • the third network element can be a location reference device.
  • the location reference device can be regarded as a special network element, which can generally be configured by the network manufacturer.
  • the network manufacturer can configure one or more of the location, transmission capability, reception capability and processing capability of the location reference device.
  • the location reference device can provide its location information to the access network device.
  • the third network element can be a reference UE, or an automated guided vehicle (AGV).
  • the third network element can also be an ordinary UE.
  • ordinary UE is relative to the location reference device. After the ordinary UE obtains its own location information through some positioning methods, it provides the location information to the access network device.
  • the access network device measures the detection reference signal from the third network element to obtain one or more measurement results.
  • the detection reference signal sent by the third network element may be SRS (sounding reference signal).
  • the access network device measures the detection reference signal sent by the third network element, and obtains one or more measurement results, and the one or more measurement results have a corresponding relationship with the position information of the third network element.
  • the third network element sends a detection reference signal at position 1, and the access network device obtains measurement result 1 by measuring the detection reference signal, and measurement result 1 corresponds to position 1.
  • the third network element sends a detection reference signal at position 2, and the access network device obtains measurement result 2 by measuring the detection reference signal, and measurement result 2 corresponds to position 2.
  • the absolute position of the third network element has not changed, but the surrounding environment of the third network element changes at different times, and the access network device measures the detection reference signal sent by the third network element at different times, and the measurement results obtained may also change.
  • the measurement result 1 obtained by the access network device at time 1 corresponds to position 1 of the third network element
  • the measurement result 2 obtained at time 2 corresponds to position 1 of the third network element.
  • the access network device measures the detection reference signals from multiple third network elements respectively to obtain multiple measurement results. That is, the multiple measurement results Each measurement result corresponds to the position of a third network element in the plurality of third network elements. Accordingly, in step 705, the plurality of third network elements respectively provide their respective position information to the access network device or the positioning device.
  • the third network element provides its own location information.
  • a third network element is used as an example for explanation.
  • One location information corresponds to one or more measurement results obtained by one or more access network devices measuring one or more detection reference signals sent by the third network element at the location corresponding to the location information.
  • the candidate training data of the AI model is the one or more measurement results and the location information of the third network element corresponding to the one or more measurement results.
  • the one or more access network devices determine that the collected candidate training data (i.e., the one or more measurement results) is valid, the one or more access network devices will provide the valid candidate training data to the positioning device respectively.
  • the location information of the third network element corresponding to the valid candidate training data may be provided by the third network element to the positioning device through at least one of the one or more access network devices, as shown in step 705a. It should be understood that 705a is an implementation of step 705.
  • the location information of the third network element may be visible to the at least one access network device, or invisible.
  • the third network element directly provides the location information of the subframe to the positioning device (not shown).
  • the positioning device obtains valid candidate training data from one or more access network devices, and location information corresponding to the valid candidate training data. It should be understood that there are multiple valid candidate training data, and there are also multiple location information of the third network element.
  • the positioning device determines the correspondence between the valid candidate training data and the location information.
  • the positioning device uses the location information as a label of the AI model to train or update the AI model, that is, the training of the new process or the training of the update process.
  • the valid candidate training data (that is, the first training data) of the AI model obtained by the positioning device includes: one or more measurement results that meet the constraint conditions in the measurement results obtained by the access network device measuring the sounding reference signal sent by the third network element, and the location information of the third network element corresponding to each measurement result.
  • the location information of the third network element is the output true value of the AI model, that is, the label.
  • the access network device determines the validity of the collected candidate training data according to the first information.
  • step 706 the access network device specifically determines the validity of the measurement results in the candidate training data.
  • the quality indicators of the measurement results may include, but are not limited to, one or more of the first path power, first path arrival delay, timing error group (TEG), average power of time domain sampling points, phase difference between antenna ports, equivalent SINR of the full band or sub-band, interference level information of the full band or sub-band, line of light (LOS) probability, inter-station synchronization error indication information, and measurement result confidence indication information.
  • the quality indicator of the tag can be the distance between the locations of different samples.
  • the quality indicators in the constraints may be SINR and the number of training data.
  • the threshold of the number of training data is N, and N may be an integer multiple of a batch or the minimum number of training data required for the AI model to converge.
  • the candidate training data of the AI model collected by the access network device also includes a label, which is location information.
  • the quality indicators in the constraints may also include quality indicators of the labels.
  • the quality indicators of the labels may be the distance between the locations of different samples, etc., which is not limited to this.
  • step 504 For the validity determination, please refer to the description of step 504, which will not be repeated here.
  • the access network device sends second information to the positioning device according to the determination result of the validity of the candidate training data, wherein the second information indicates the determination result of the validity.
  • the second information may be included in an interface message between the access network device and the positioning device.
  • the access network device sends an interface message to the positioning device, and the interface message includes the second information.
  • the positioning device determines whether the candidate training data collected by the access network device is valid according to the second information.
  • the second information indicates that the candidate training data collected by the access network device is valid.
  • the positioning device obtains the valid candidate training data (i.e., the first training data) collected by the access network device.
  • the first training data specifically includes one or more measurement results that meet the constraint conditions and the location information of the third network element corresponding to each of the one or more measurement results. Further, the positioning device trains or updates the AI model according to the valid candidate training data, as in step 709.
  • the positioning device trains the AI model to obtain the AI model or updates the AI model.
  • the second information indicates that the candidate training data collected by the access network device is invalid.
  • the positioning device maintains the original AI model, or switches to a non-AI model, such as step 710.
  • the positioning device performs uplink positioning based on the original AI model or switches to a non-AI model.
  • the second information indicates that the candidate training data collected by the access network device is invalid.
  • the positioning device determines to re-collect the training data of the AI model. In this case, steps 711 and 712 are also included.
  • the positioning device determines to collect training data again.
  • the positioning device sends third information to the access network device, and the third information instructs the access network device to re-collect training data for the AI model.
  • the third information further indicates the maximum number of validity determinations k, where k is a positive integer.
  • the maximum number of validity determinations k may also be indicated by the first information, and reference may be made to the relevant description in the process shown in FIG3 , which will not be described in detail.
  • the maximum number of validity determinations k can be configured by the positioning device according to the urgency of the training data requirements of the AI model, which is similar to that in application scenario 1.
  • the urgency judgment criterion can be determined based on the error of the estimation result of the current AI model for the position of the third network element or the time interval from the last update time of the AI model to the present. For example, if the error of the estimation result of the position of the third network element by the positioning device based on the current AI model is large, for example, greater than or equal to a certain set threshold, it can be determined as urgent.
  • the maximum number of validity determinations k can be set larger; conversely, if the error of the estimation result of the position of the third network element based on the current AI model is small, for example, less than the set threshold, it can be determined as not urgent. In this case, the maximum number of validity determinations k can be set smaller.
  • the judgment of the error of the estimation result of the AI model is achieved by splitting the training data collected by the last AI model training into a training set and a verification set. Since the error of the training set is already very low, the error of the estimation result of the verification set is used as a criterion for judging whether the AI model is seriously invalid. In addition, it can also be set based on the time interval from the last update time of the AI model to the present. Please refer to the explanation in application scenario 1 and will not be repeated here.
  • the positioning device may instruct the access network device to update the air interface transmission configuration when the access network device collects training data, as in step 713.
  • the access network device sends air interface transmission configuration information (such as air interface transmission configuration information #2) to the third network element, where the air interface transmission configuration information indicates an updated air interface transmission configuration.
  • air interface transmission configuration information such as air interface transmission configuration information #2
  • the update of the air interface transmission configuration in step 713 is an update relative to the air interface transmission configuration in step 703.
  • the update includes but is not limited to: increasing the transmit power of the sounding reference signal, increasing the frequency domain density of the sounding reference signal, etc.
  • the purpose of updating the air interface transmission configuration is that the access network device attempts to collect candidate training data that meets the constraint conditions and provide it to the positioning device.
  • the access network device re-collects the training data of the AI model.
  • the access network device is an example of a network element for collecting training data
  • the positioning device is an example of a network element for training an AI model.
  • FIG 8 is a schematic diagram of the technical solution provided by this application in a downlink positioning scenario based on an AI model.
  • AI model inference is deployed on the UE side, but the training of the AI model is deployed on the positioning device on the network side, such as an LMF network element.
  • the AI model deployed on the positioning device takes the corresponding channel response obtained by the UE measuring the reference signal as input and the position of the UE as output.
  • the reference signal can be a positioning reference signal, which can be sent to the UE by one or more base stations (BS).
  • BS base stations
  • the positioning device determines that it is necessary to collect training data for the AI model.
  • the training data of the AI model may come from a UE's measurement of multiple reference signals, or each of multiple UEs' measurement of one or more reference signals.
  • the multiple reference signals may come from one or more access network devices. This embodiment is described from the perspective of communication between the positioning device and the one UE or one of the multiple UEs.
  • the positioning device sends first information to the UE, where the first information is used to determine the validity of the candidate training data collected by the UE.
  • the determination result of the validity may be valid or invalid.
  • the positioning device sends the first information to the UE through the access network device, such as steps 802a and 802b shown in Figure 9.
  • the positioning device may also directly send the first information to the UE through the interface between the positioning device and the UE.
  • the positioning device sends information #1 to the access network device, and information #1 indicates that the access network device sends a positioning reference signal to the UE.
  • information #1 may also be the first information.
  • the access network device sends a positioning reference signal to the UE based on the triggering of information #1.
  • the implementation shown in Figure 9 is only an example.
  • the access network device sends a positioning reference signal to the UE.
  • the UE measures a positioning reference signal from an access network device, or the UE measures a positioning reference signal from the access network device and other access network devices, and obtains candidate training data, specifically one or more measurement results of the positioning reference signal, and the UE location information corresponding to the one or more measurement results.
  • the access network device sends the PRS to the UE, it also sends the air interface transmission configuration information of the PRS to the UE to indicate the air interface transmission configuration of the PRS.
  • the positioning reference signal sent by the access network device to the UE may be a PRS.
  • the candidate training data is one or more measurement results obtained by the UE measuring the PRS and the location information of the UE corresponding to the one or more measurement results.
  • the UE determines the validity of the collected candidate training data according to the first information.
  • the UE determines the validity of the collected candidate training data according to the constraint condition indicated by the first information, specifically determining the validity of the one or more measurement results. Similar to step 504, it can be understood with reference to step 504, and detailed description is omitted here. In addition, for examples of quality indicators included in the constraint condition in the downlink positioning scenario, reference can be made to the description in the uplink positioning scenario, which will not be repeated here.
  • the UE sends second information to the positioning device, where the second information indicates a validity determination result.
  • the second information indicates that the candidate training data collected by the UE is valid.
  • the second information may be the first training data among the candidate training data collected by the UE.
  • the first training data includes the location information of the UE.
  • the first training data is specifically the measurement result that meets the constraint condition and the location information of the UE corresponding thereto.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the UE may directly send the second information to the positioning device through the interface between the UE and the positioning device, as shown in FIG9 .
  • the UE may also send the second information to the access network device, and the access network device then sends the second information to the positioning device.
  • the UE sends part of the information contained in the second information to the access network device, such as the measurement results (i.e., valid measurement results) that meet the constraint conditions contained in the first training data, and sends its own location information to the positioning device.
  • the access network device then sends the measurement results that meet the constraint conditions to the positioning device.
  • the positioning device thereby obtains the first training data, wherein the first training data includes valid measurement results and their corresponding UE location information, without limitation.
  • the positioning device determines whether the candidate training data collected by the UE is valid according to the second information.
  • the second information indicates that the candidate training data collected by the UE is valid, in which case the second information may include first training data, wherein the first training data includes the location information of the UE. Further, the positioning device trains or updates the AI model according to the first training data, as in step 807.
  • the positioning device trains the AI model to obtain the AI model or updates the AI model.
  • the access network device maintains the original AI model beam management or switches to a non-AI model for downlink positioning, such as step 808.
  • the positioning device maintains the original AI model or switches to a non-AI model for positioning.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the positioning device determines to collect the training data again, such as step 809 .
  • the positioning device determines to collect training data again.
  • the positioning device sends third information to the UE, and the third information instructs the UE to re-collect training data of the AI model.
  • the third information indication may also indicate a maximum number k of validity determinations, or the first information indication may indicate a maximum number k of validity determinations.
  • the positioning device may instruct the access network device to update the air interface transmission configuration.
  • the positioning device sends information #2 to the access network device, and information #2 is used to instruct the access network device to recollect the training data.
  • the access network device sends the air interface transmission configuration information corresponding to the updated air interface transmission configuration to the UE, as shown in step 811.
  • the access network device sends air interface transmission configuration information to the UE, where the air interface transmission configuration information indicates an updated air interface transmission configuration.
  • the UE Based on the updated air interface transmission configuration, the UE measures the positioning reference signal sent by the access network device to re-collect training data for the AI model.
  • the UE collects training data for the AI model again.
  • the UE in this embodiment may be a location reference device or a common UE, without limitation.
  • the UE may refer to the description in step 703 and will not repeat them here.
  • the UE in this embodiment is an example of a network element for collecting training data
  • the positioning device is an example of a network element for training an AI model.
  • the training of the AI model in application scenario 3 is deployed on the access network device side.
  • the access network device needs to obtain the training data (for example, one or more measurement results of the reference signal) obtained by measuring the reference signal on the UE side, and use the training data obtained from the UE side for training or updating the AI model.
  • the label of the AI model is the information of the reference signal corresponding to the optimal K measurement results.
  • the information of the reference signal corresponding to the optimal K measurement results can also be replaced by the information of the K beams corresponding to the optimal K measurement results, exemplarily K beam IDs.
  • the 5G system introduces high frequency bands above 6GHz for data communication. Compared with the medium and low frequency bands below 6GHz, the continuous available bandwidth of the high frequency band spectrum is larger and the center frequency is higher, so a higher transmission rate and system capacity can be obtained.
  • the weak penetration ability of high-frequency signals such as millimeter waves
  • the strong path fading effect the propagation distance of high-frequency signals is limited and the coverage capability is worrying.
  • high-frequency communication systems usually use a large number of antennas for beamforming, so that considerable beam gain can be obtained to compensate for the limited propagation distance caused by high-frequency propagation characteristics.
  • the base station needs to obtain accurate channel information from the terminal.
  • the transmitter has a total of S antennas and the receiver has R antennas, which can be linear antennas or array antennas.
  • the transmitter and receiver multiply their antennas by different precoding weights to precode the transmitted signal, so that the transmitted signal has a beamforming effect.
  • Y VHWX+N
  • the receiving precoding matrix of the receiving end is V, and the channel response is H.
  • the transmitting precoding matrix of the transmitting end is W
  • the transmitting signal is X
  • the noise is N.
  • the signal received by the receiving end is Y.
  • WX The signal obtained after the transmitting signal X is precoded by W is WX, which is the final transmitting signal of the transmitting end.
  • WX has a beamforming effect in space. According to the difference of the information carried on X, WX can be divided into a reference signal and a data signal.
  • the reference signal is generally sent during the beam management process. Possible reference signals include SSB, CSI-RS, SRS, phase tracking reference signal (PTRS), demodulation reference signal (DMRS), etc.
  • each precoding matrix W can only cover a certain angle range in space, corresponding to a shaped beam, it is necessary to design multiple precoding matrices W to ensure a good signal coverage effect.
  • Multiple precoding matrices W with different pointing angles constitute a codebook.
  • Both the transmitter and the receiver maintain their own codebooks. In the beam management process, the transmitter and the receiver achieve angle alignment between the transmitter and the receiver by traversing and scanning their own codebooks. For example, there are 64 precoding matrices in the transmitter's codebook, corresponding to 64 shaped beams respectively.
  • the receiver's codebook has 4 precoding matrices, corresponding to 4 shaped beams respectively, so a total of 256 (64*4) scans are required to determine a pair of optimal shaped beams for the transmitter and receiver, and the scanning overhead and delay are very large.
  • any transmitting shaped beam may constitute a transceiver beam pair with it.
  • the process of determining the optimal transceiver beam pair can be decomposed into performing a beam scan on the transmitter to determine the matching optimal transmit beam for a certain receiving beam, and then repeating this process for the remaining R-1 receiving beams to determine the globally optimal transceiver beam pair.
  • any receiving beam may constitute a transceiver beam pair with it. Therefore, the process of determining the optimal transceiver beam pair can be decomposed into performing a beam scan on the receiver to determine the matching optimal transmit beam for a certain transmitting beam, and then repeating this process for the remaining S-1 transmitting beams to determine the globally optimal transceiver beam pair. Therefore, the following article takes the transmitting beam scanning as an example for explanation.
  • the traditional scheme needs to scan all 64 beams to determine the optimal beam, but the AI-assisted (i.e., AI model-based) sparse beam scanning scheme only needs to scan part of the beams in the codebook, such as the second value of the fill mark in Figure 10, such as 16 beams.
  • the transmitter uses the precoding matrix in the sparse beam pattern for precoding and sends a reference signal.
  • the receiver inputs the measurement result of the reference signal into the neural network for AI beam prediction, and the neural network outputs the index of K beams, also called the index of Top-K beams.
  • K beams are K of all 64 shaped beams in the codebook, and are not limited to the K of the shaped beams contained in the sparse beam pattern.
  • the receiver feeds back the index of the Top-K beam to the transmitter.
  • the transmitter only scans the K beamforming beams and transmits the reference signals after beamforming.
  • the receiver uses an energy detection method to measure the energy of the K reference signals and selects the one with the strongest energy as the optimal beam.
  • the transmitting end is an access network device, such as a base station.
  • the receiving end is a UE.
  • the UE measures a reference signal from the access network device, obtains multiple measurement results, determines a TOP-K beam index, and feeds back the multiple measurement results and the TOP-K beam index to the access network device.
  • the access network device determines that it is necessary to collect training data for the AI model.
  • the access network device sends first information to the UE, where the first information is used to determine the validity of the candidate training data collected by the UE.
  • the determination result of the validity may be valid or invalid.
  • the access network device sends multiple reference signals to the UE.
  • the UE measures multiple reference signals from the access network device to obtain multiple measurement results, namely candidate training data.
  • the multiple reference signals correspond to the aforementioned first value, such as 64, shaped beams.
  • the reference signal may be a CSI-RS and/or an SSB.
  • the CSI-RS and/or the SSB are used for the UE to perform channel measurement.
  • the UE determines the validity of the collected candidate training data according to the first information.
  • the UE sends second information to the access network device, where the second information indicates a validity determination result.
  • the second information indicates that the candidate training data collected by the UE is valid.
  • the second information may be the first training data, wherein the first training data includes information about reference signals corresponding to the K best measurement results among the one or more measurement results, and K is an integer greater than or equal to 1.
  • the information about the reference signals corresponding to the K best measurement results can be understood as the index of the TOP-K beam in Figure 10.
  • the second information may be the first training data, wherein the first training data includes L measurement results among the multiple measurement results, and information about L reference signals corresponding to the L measurement results. Among them, the L measurement results are valid measurement results, that is, measurement results that meet the constraint conditions.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the reference signal can also be replaced by "beam” without limitation.
  • the access network device determines whether the candidate training data collected by the UE is valid according to the second information.
  • the second information indicates that the candidate training data collected by the UE is valid, in which case the access network device obtains the first training data from the UE. Further, the access network device trains or updates the AI model according to the first training data, as in step 907.
  • the access network device trains the AI model according to the first training data to obtain the AI model or update the AI model.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the access network device maintains the original AI model beam management or switches to a non-AI model for beam management, such as step 908.
  • the access network device maintains the original AI model unchanged, or switches to a non-AI model for beam management.
  • the second information indicates that the candidate training data collected by the UE is invalid.
  • the access network device determines to collect the training data again, such as step 909 .
  • the access network device determines to collect training data again.
  • the access network device sends third information to the UE, and the third information instructs the UE to re-collect training data of the AI model.
  • the third information further indicates a maximum number k of validity determinations.
  • the maximum number k of validity determinations may also be indicated by the first information, without limitation.
  • the access network device may update the air interface transmission configuration. Accordingly, the UE re-collects the candidate training data of the AI model based on the updated air interface transmission configuration.
  • the access network device sends air interface transmission configuration information to the UE, where the air interface transmission configuration information indicates an updated air interface transmission configuration.
  • UE collects training data for the AI model again.
  • the UE in the process of recollecting candidate training data, the UE is limited by the maximum number of validity determinations k.
  • the maximum number of validity determinations k can be determined by the access network device according to the urgency of training data collection.
  • the urgency judgment criterion can refer to the time since the last update of the AI model, or the access network device is set according to the current AI model for the error of the estimated result of the optimal measurement result (or the optimal beam).
  • the judgment of the error of the estimated result of the AI model is achieved by splitting the training data collected by the last AI model training into a training set and a validation set.
  • the error of the training set is already very low, the error of the estimated result of the validation set is used as a criterion for judging whether the AI model is seriously ineffective. For example, if the access network device predicts the information of the reference signal corresponding to the optimal measurement result when the UE receives the reference signal according to the current AI model, the error is large, for example, greater than or equal to a certain set threshold, it can be determined to be urgent. In this case, the maximum number of validity determinations k can be set larger; on the contrary, if the error of the predicted result of the information of the reference signal corresponding to the optimal measurement result when the UE receives the reference signal based on the current AI model is small, for example, less than the set threshold, it can be determined to be not urgent. In this case, the maximum number of validity determinations k can be set smaller. In addition, it can also be set according to the time interval from the last update time of the AI model to the present. Please refer to the explanation in application scenario 1 and do not elaborate on it.
  • the present application provides a communication device 1000 .
  • the communication device 1000 includes a processing module 1001 and a communication module 1002.
  • the communication device 1000 can be a terminal device, or a communication device applied to a terminal device or used in combination with a terminal device, which can implement a communication method executed by the terminal device side, such as a chip or a circuit; or, the communication device 1000 can be a network device, or a communication device applied to a network device side or used in combination with a network device side, which can implement a communication method executed by the network device side, such as a chip or a circuit.
  • the network device side can be, for example, an access network device or a positioning device in the method embodiment of the present application.
  • the communication module may also be referred to as a transceiver module, a transceiver, a transceiver, or a transceiver device, etc.
  • the processing module may also be referred to as a processor, a processing board, a processing unit, or a processing device, etc.
  • the communication module is used to perform the sending operation and the receiving operation on the terminal device side or the network device side in the above method, and the device used to implement the receiving function in the communication module may be regarded as a receiving unit, and the device used to implement the sending function in the communication module may be regarded as a sending unit, that is, the communication module includes a receiving unit and a sending unit.
  • the processing module 1001 can be used to implement the processing function of the terminal device in the embodiments described in Figures 3 to 11, and the communication module 1002 can be used to implement the transceiver function of the terminal device in the embodiments described in Figures 3 to 11.
  • the communication device can also be understood by referring to the third aspect in the content of the invention and the possible designs in the third aspect.
  • the processing module 1001 can be used to implement the network in each embodiment described in FIG. 3 to FIG. 11.
  • the communication module 1002 can be used to implement the transceiver function of the network device in each embodiment described in Figures 3 to 11.
  • the communication device can also be understood by referring to the fourth aspect of the invention and the possible design in the fourth aspect.
  • first network element or the second network element shown in Figure 3 is specifically a terminal device or a network device (such as an access network device or a positioning device), which has been described in detail in the aforementioned method embodiments for various application scenarios. You can refer to the specific embodiments to understand that the first network element is a terminal device or a network device, which will not be repeated here.
  • the aforementioned communication module and/or processing module can be implemented through a virtual module, for example, the processing module can be implemented through a software function unit or a virtual device, and the communication module can be implemented through a software function or a virtual device.
  • the processing module or the communication module can also be implemented through a physical device, for example, if the device is implemented using a chip/chip circuit, the communication module can be an input-output circuit and/or a communication interface, performing input operations (corresponding to the aforementioned receiving operations) and output operations (corresponding to the aforementioned sending operations); the processing module is an integrated processor or microprocessor or integrated circuit.
  • each functional module in each example of this application may be integrated into one processor, or may exist physically separately, or two or more modules may be integrated into one module.
  • the above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules.
  • the present application also provides a communication device 1100.
  • the communication device 1100 may be a chip or a chip system.
  • the chip system may be composed of a chip, or may include a chip and other discrete devices.
  • the communication device 1100 can be used to implement the functions of any network element in the communication system described in the above examples.
  • the communication device 1100 may include at least one processor 1110.
  • the processor 1110 is coupled to a memory, and the memory may be located within the device, or the memory may be integrated with the processor, or the memory may be located outside the device.
  • the communication device 1100 may also include at least one memory 1120.
  • the memory 1120 stores the necessary computer programs, computer programs or instructions and/or data for implementing any of the above examples; the processor 1110 may execute the computer program stored in the memory 1120 to complete the method in any of the above examples.
  • the communication device 1100 may also include a communication interface 1130, and the communication device 1100 may exchange information with other devices through the communication interface 1130.
  • the communication interface 1130 may be a transceiver, a circuit, a bus, a module, a pin, or other types of communication interfaces.
  • the communication interface 1130 in the device 1100 may also be an input-output circuit, which may input information (or receive information) and output information (or send information)
  • the processor may be an integrated processor or a microprocessor or an integrated circuit or a logic circuit, and the processor may determine the output information based on the input information.
  • the coupling in this application is an indirect coupling or communication connection between devices, units or modules, which can be electrical, mechanical or other forms, and is used for information exchange between devices, units or modules.
  • the processor 1110 may cooperate with the memory 1120 and the communication interface 1130.
  • the specific connection medium between the above-mentioned processor 1110, memory 1120 and communication interface 1130 is not limited in this application.
  • the processor 1110, the memory 1120, and the communication interface 1130 are interconnected via a bus 1140.
  • the bus 1140 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus.
  • PCI peripheral component interconnect
  • EISA extended industry standard architecture
  • the bus may be divided into an address bus, a data bus, a control bus, and the like.
  • FIG13 is represented by only one thick line, but it does not mean that there is only one bus or one type of bus.
  • the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or execute the methods, steps, and logic block diagrams disclosed in this application.
  • a general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the method disclosed in this application may be directly embodied as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.
  • the memory may be a non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), etc., or a volatile memory (volatile memory), such as a random-access memory (RAM).
  • the memory is any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.
  • the memory in the present application may also be a circuit or any other device that can realize a storage function, used to store program instructions and/or data.
  • the communication device 1100 can be applied to a network device side, such as an access device in an embodiment of the present application.
  • Network device or positioning device Specifically, the communication device 1100 can be a network device, or it can be a device that can support the network device to implement the corresponding functions on the network device side in any of the above-mentioned examples.
  • the memory 1120 stores computer programs (or instructions) and/or data that implement the functions on the network device side in any of the above-mentioned examples.
  • the processor 1110 can execute the computer program stored in the memory 1120 to complete the method executed on the network device side in any of the above-mentioned examples.
  • the communication interface in the communication device 1100 can be used to interact with a terminal device, send information to a terminal device, or receive information from a terminal device; in addition, optionally, the communication interface in the communication device 1000 can also be used to interact with a core network device, such as interacting with a positioning device (such as an LMF network element), sending information to a positioning device, or receiving information from a positioning device.
  • a positioning device such as an LMF network element
  • the communication device 1100 can be applied to a terminal device.
  • the communication device 1100 can be a terminal device, or a device that can support a terminal device and implement the functions of the terminal device in any of the above-mentioned examples.
  • the memory 1120 stores a computer program (or instruction) and/or data that implements the functions of the terminal device in any of the above-mentioned examples.
  • the processor 1110 can execute the computer program stored in the memory 1120 to complete the method executed by the terminal device in any of the above-mentioned examples.
  • the communication interface in the communication device 1100 can be used to interact with a network device side (for example, an access network device), send information to the network device side, or receive information from the access network device.
  • a network device side for example, an access network device
  • the communication device 1100 provided in this example can be applied to a network device side (such as an access network device or a positioning device) to complete the method executed by the above network device side, or applied to a terminal device to complete the method executed by the terminal device, the technical effects that can be obtained can refer to the description in the above method embodiment, and will not be repeated here.
  • a network device side such as an access network device or a positioning device
  • the present application provides a communication system, including a network device and a terminal device.
  • the communication system includes an access network device and a terminal device.
  • the communication system includes a positioning device, an access network device, and a terminal device.
  • the access network device and the terminal device, or the positioning device, the access network device, and the terminal device can implement the communication method provided in the examples shown in Figures 3 to 11.
  • the technical solution provided in this application can be implemented in whole or in part by software, hardware, firmware or any combination thereof.
  • software When implemented by software, it can be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions.
  • the computer can be a general-purpose computer, a special-purpose computer, a computer network, a terminal device, an access network device or other programmable device.
  • the computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium.
  • the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.
  • the computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated.
  • the available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a digital video disc (DVD)), or a semiconductor medium, etc.
  • the examples may reference each other, for example, the methods and/or terms between method embodiments may reference each other, for example, the functions and/or terms between device embodiments may reference each other, for example, the functions and/or terms between device examples and method examples may reference each other.
  • a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program and/or a computer.
  • applications and computing devices running on a computing device can be components.
  • One or more components may reside in a process and/or an execution thread, and a component may be located on a computer and/or distributed between two or more computers.
  • these components may be executed from various computer-readable media having various data structures stored thereon.
  • Components may, for example, communicate through local and/or remote processes according to signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system and/or a network, such as the Internet interacting with other systems through signals).
  • signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system and/or a network, such as the Internet interacting with other systems through signals).
  • the disclosed systems, devices and methods can be implemented in other ways.
  • the device embodiments described above are only schematic.
  • the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.
  • Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
  • the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
  • the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art.
  • the computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Quality & Reliability (AREA)
  • Electromagnetism (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请提供一种AI模型训练中用于获取训练数据的方法，可应用于AI模型的训练网元(第二网元)和训练数据的收集网元(第一网元)在逻辑上分离部署的场景。第一网元接收来自于第二网元的第一信息，第一信息用于第一网元所收集的候选训练数据的有效性的判定。第一网元收集AI模型的候选训练数据，并根据第一信息判定候选训练数据的有效性。在判定候选训练数据有效的情况下，第一网元向第二网元发送有效的候选训练数据，而不发送无效的候选训练数据；在收集的候选训练数据中不包含有效的候选训练数据的情况下，第一网元向第二网元指示此次收集的训练数据无效，不向第二网元发送此次收集的候选训练数据，可以减少空口资源浪费。

Description

AI模型训练中用于获取训练数据的方法以及通信装置
本申请要求于2022年09月29日提交中国国家知识产权局、申请号为“202211203052.X”、申请名称为“AI模型训练中用于获取训练数据的方法以及通信装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及机器学习领域，更具体地，涉及一种AI模型训练中用于获取训练数据的方法以及通信装置。
背景技术
在AI模型应用于空口技术的一些应用场景中时，AI模型的训练和AI模型的训练数据的收集可能部署在不同的网元。在此现状之下，AI模型的训练或更新需要AI模型的训练网元和训练数据的收集网元之间进行训练数据(例如，参考信号的测量结果和/或标签)的交互。
在现有的方案中，AI模型的训练网元和训练数据的收集网元之间的训练数据的交互通常是固定周期发送或持续存在的。而这种信息交互方式容易造成空口资源的浪费。
发明内容
本申请提供一种AI模型训练中用于获取训练数据的方法和通信装置，以期减少空口资源的浪费。
第一方面，提供了一种AI模型训练中用于获取训练数据的方法，可以应用于训练数据的收集网元，例如终端设备或接入网设备，该方法包括：
第一网元接收来自于第二网元的第一信息，所述第一信息用于所述第一网元收集的候选训练数据的有效性的判定，所述有效性的判定结果包括有效或无效；
所述第一网元收集所述AI模型的候选训练数据；
所述第一网元根据所述候选训练数据和所述第一信息，向第二网元发送第二信息，所述第二信息指示所述有效性的判定结果。
在本申请的技术方案中，第一网元为收集AI模型的训练数据的网元，第二网元为训练AI模型的网元。第二网元需要第一网元收集AI模型的训练数据时，向第一网元发送第一信息，第一信息用于指示第一网元收集AI模型的训练数据，同时也用于第一网元所收集的候选训练数据的有效性的判定(也简称为有效性判定)。第一网元收集AI模型的候选训练数据，并根据第一信息判定所收集的候选训练数据的有效性。之后，第一网元向第二网元发送第二信息，以指示该有效性的判定结果。基于该技术方案，第一网元在完成一次候选训练数据的收集之后，会对收集到的候选训练数据会进行有效性的判定。只有有效的候选训练数据才作为训练数据由第一网元提供给第二网元使用，而不是将收集的数据不作任何筛选地提供给第二网元。可以减少收集到的无效的候选训练数据的传输，从而降低空口资源的浪费。
结合第一方面，在第一方面的某些实现方式中，第一信息用于所述有效性判定所使用的约束条件的确定。可选的，所述约束条件可以包括如下一项或多项：
质量指标的门限和所述质量指标的判定准则；或，
符合质量指标的判定准则的训练数据的数量门限和所述训练数据的数量的判定准则；或，
单次有效性判定对应的候选训练数据收集的最大时长。
结合第一方面，在第一方面的某些实现方式中，所述第一信息指示如下一项或多项：
质量指标的门限；
质量指标的判定准则；
符合质量指标的判定准则的训练数据的数量门限；
符合质量指标的判定准则的训练数据的数量的判定准则；或，
单次有效性判定对应的候选训练数据收集的最大时长。
可以理解的是，如上信息中未被第一信息指示的部分，可以由协议预定义。
可选的，第一信息指示如上信息中的部分，可以包括显式指示如上信息中的部分中的一项或多项，或者，隐式指示如上信息中的部分中的一项或多项。显式指示可以包括：第一信息包括如上信息中其所显式指示的部分中的一项或多项。隐式指示可以包括：第一信息包括和如上信息中其所隐式指示的部分中的一项或多项对应的其他信息。可选的，该其他信息可以包括与如上信息中其所隐式指示的部分中的一项或多项具有对应关系的索引。其中，该其他信息可以包括一个或多个信息，其中，多个信息各自指示如上信息中被隐式指示的部分中各项信息的部分。
可选的，如上对应关系可以是协议预定义的，或是，预存储，预先配置的。其中，预先配置，可以采用无线资源控制(radio resource control，RRC)信令，配置多个索引与如上信息中一项或多项的组合的多个值的对应关系。
可选的，如上第一信息可以携带在控制信息，如下行控制信息(downlink control information，DCI)中。
可选的，如上质量指标可以包括一个或多个质量指标，其各自具有对应的门限及判定准则。比如，质量指标可以包括对参考信号的测量结果的质量指标，或，标签的质量指标中的一项或多项。其中，标签用于作为AI模型训练的比较真值。比如，标签可以包括位置信息，波束图样(pattern)，信道测量结果等中的一项或多项。
可选的，所述质量指标的门限可以包括如上符合质量指标的判定准则的训练数据的数量门限，所述质量指标的判定准则可以包括符合质量指标的判定准则的训练数据的数量的判定准则。
在本申请中，第一信息可以用于约束条件的确定，具体可以有多种实现方式，下面举几个例子进行说明。
在一个示例中，第一信息指示一个或多个质量指标的门限，该一个或多个质量指标的判定准则是由协议预定义的。例如，质量指标包括训练数据的信干噪比(signal to interference plus noise ratio,SINR)和训练数据的数量，SINR的判定准则为：SINR大于或等于门限Q；所述训练数据的数量的判定准则为：所述训练数据的数量大于或等于门限N。示例性地，第一信息指示Q和N，SINR的判定准则和所述训练数据的数量的判定准则均是由协议预定义的。
在该示例中，通过协议预定义质量指标的判定准则，可以节省指示开销。
在另一个示例中，第一信息指示一个或多个质量指标的门限，以及该一个或多个质量指标的判定准则。例如，质量指标包括训练数据的SINR和训练数据的数量，SINR的判定准则为：SINR大于或等于门限Q，所述训练数据的数量的判定准则为：所述训练数据的数量大于或等于门限N。示例性地，第一信息指示Q和N，此外，第一信息包含信息域，该信息域用于指示SINR的判定准则和所述训练数据的数量的判定准则。例如，若信息域的取值为1，表示“SINR大于或等于Q，且所述训练数据的数量大于或等于N”；若信息域的取值为0，表示“SINR大于Q，且所述训练数据的数量大于N”。
在该示例中，第一信息指示约束条件质量指标的门限以及该质量指标的判定准则，使得第二网元可以根据对训练数据的需求的变化，适应性更新约束条件，适用于约束条件变化较为频繁的场景，可以提升AI模型对于不同应用场景的适应性，并且提升了在不同应用场景下收集到符合要求的训练数据的概率。
在再一个示例中，第一信息指示部分质量指标的门限，另一部分质量指标的门限以及这些质量指标的判定准则由协议预定义。例如，质量指标包括训练数据的SINR和训练数据的数量，SINR的判定准则为：训练数据的SINR大于或等于门限Q，所述训练数据的数量的判定准则为：所述训练数据的数量大于或等于门限N。示例性地，第一信息指示Q，而所述训练数据的数量的门限N，以及SINR的判定准则和所述训练数据的数量的判定准则可以由协议预定义。
在该示例中，可以将应用场景中变化周期较长的质量指标的门限及其判定准则通过协议预定义，以节省信令开销；而将变化较为频繁的质量指标的门限及其判定准则通过第一信息来指示，可以保证对于所需训练数据的要求的灵活调整。该示例可以兼顾信令开销和约束条件更新的灵活性。
在再一个示例中，第一信息指示部分质量指标的门限和一个索引信息，该索引信息用于确定所述 部分质量指标的判定准则，以及约束条件中其它质量指标的门限以及所述其它质量指标的判定准则。示例性地，第一信息指示SINR的门限Q和index 0,其中，index 0表示：训练数据的数量的门限为N，SINR的判定准则为：训练数据的SINR大于或等于Q，且，训练数据的数量的判定准则为：训练数据的数量至少为N个。可选地，index 0为的多个应用场景中的一个应用场景下的索引值，例如，该多个应用场景包括但不限于CSI预测、上行定位、下行定位或波束管理，index 0为波束管理场景对应的一个或多个index中的一个index。可选地，index 0为某个应用场景下的索引值，例如，上行定位场景下对应多个index，index 0为该多个index中的一个。
在再一个示例中，第一信息指示一个索引信息，该索引信息用于确定一个或多个质量指标的门限，以及所述一个或多个质量指标的判定准则。示例性地，第一信息指示index 0,其中，index 0表示：训练数据SINR的门限为Q，训练数据的数量的门限为N，SINR的判定准则为：训练数据的SINR大于或等于Q，且，训练数据的数量的判定准则为：训练数据的数量至少为N个。可选地，index 0为的多个应用场景中的一个应用场景下的索引值，例如，该多个应用场景包括但不限于CSI预测、上行定位、下行定位或波束管理，index 0为波束管理场景对应的一个或多个index中的一个index。可选地，index0为某个应用场景下的索引值，例如，上行定位场景下对应多个index，index 0为该多个index中的一个。
在上述后两个示例中，索引信息和质量指标的门限和/或质量指标的判定准则的对应关系由协议预定义也仅是作为示例，也可以为其它可实现的方式，包括但不限于为预存储或预先配置等。
以上是关于第一信息用于确定约束条件的示例说明，本申请不限定于上述示例。
可选地，结合第一方面，在第一方面的某些实现方式中，所述第一信息所指示的质量指标包括AI模型的标签的质量指标。
可选地，AI模型的训练数据还包括标签。作为一个示例，在上行定位或下行定位的应用场景下，该标签为位置信息。可选地，约束条件中的质量指标还可以包括标签的质量指标，例如，标签的质量指标可以包括不同样本的位置之间距离的门限等。可选地，上述第一信息所指示的质量指标还包括AI模型的标签的质量指标。
结合第一方面，在第一方面的某些实现方式中，所述第二信息包括第一训练数据且所述第二信息指示所述第一网元收集的所述候选训练数据有效，所述第一训练数据为所述候选训练数据中的有效数据。
在该实现方式中，第一网元根据第一信息对收集到的候选训练数据进行有效性判定之后，若确定此次收集是有效的，则第一网元向第二网元发送第二信息，该第二信息可以为有效的候选训练数据(即，第一训练数据)，而不发送无效的候选训练数据，由此可以降低空口资源的浪费。
在本申请中，有效的候选训练数据会被提供给第二网元用于AI模型的训练或更新，也即有效的候选训练数据实际上即成为训练数据。而无效的候选训练数据也即不符合约束条件的候选训练数据。
此外，由于第一网元不会向第二网元发送无效的候选训练数据，因而第二网元不会接收到无效或不合格的训练数据，由此避免了对整个训练数据集造成污染。同时，也避免了对第二网元训练AI模型带来不利影响，例如，利用无效的候选训练数据进行AI模型训练导致的AI性能增益评估不准、AI模型过拟合、泛化能力弱以及场景适应能力差等问题。
结合第一方面，在第一方面的某些实现方式中，所述第二信息指示所述第一网元收集的所述候选训练数据无效。
在该实现方式中，第一网元根据第一信息对收集到的候选训练数据进行有效性判定之后，若确定此次收集是无效的，则第一网元向第二网元发送第二信息，该第二信息仅指示此次收集的候选训练数据无效，而不向第二网元提供收集到的候选训练数据，由此可以降低空口资源的浪费。此外，由于第二网元不会接收到无效或不合格的候选训练数据，因此避免了对整个训练数据集造成污染；同时，也避免了对第二网元训练AI模型带来不利影响，例如，利用无效的候选训练数据进行AI模型训练导致的AI性能增益评估不准、AI模型过拟合、泛化能力弱以及场景适应能力差等问题。
结合第一方面，在第一方面的某些实现方式中，所述第一信息用于所述第一网元收集的候选训练数据的有效性的判定的约束条件的确定。
在该实现方式中，第二网元在通过第一信息指示第一网元收集AI模型的训练数据的情况下，同时 第一信息也用于第一网元确定待收集的训练数据应满足的约束条件，以便第一网元在收集到候选训练数据之后进行筛选(即有效性判定)，为第一网元判定所收集到的候选训练数据是否有效提供了的依据。
结合第一方面，在第一方面的某些实现方式中，该方法还包括：
若所述第一网元确定所述候选训练数据中包含满足所述约束条件的第一训练数据，所述第一网元确定所述候选训练数据有效；或者，
若所述第一网元确定所述候选训练数据中不包含满足所述约束条件的第一训练数据，所述第一网元确定所述候选训练数据无效。
在该实现方式中，收集到的候选训练数据中满足约束条件的候选训练数据的集合称为第一训练数据；而在不存在满足约束条件的候选训练数据的情况下，代表此次收集无效。
结合第一方面，在第一方面的某些实现方式中，在所述第一网元收集的所述候选训练数据无效的情况下，所述方法还包括：
所述第一网元接收来自于所述第二网元的第三信息，所述第三信息指示所述第一网元重新收集所述AI模型的候选训练数据。
在该实现方式中，在一次收集无效之后进行重新收集，可以将之前收集的无效的候选训练数据与重新采集的候选训练数据一起进行有效性判定，以提高获取到符合要求的候选训练数据(也即获取到训练数据)的概率。此外，还因为可以在重新收集时更新参考信号的空口传输配置，提高了获取到高质量候选训练数据的可能性，使得收集到合格训练数据的概率提高。
结合第一方面，在第一方面的某些实现方式中，所述方法还包括：
所述第一网元确定空口传输配置信息，所述空口传输配置信息对应更新的空口传输配置，所述空口传输配置信息指示所述第一网元基于所述更新的空口传输配置收集所述AI模型的候选训练数据；
其中，所述更新的空口传输配置信息包括如下一项或多项的更新：
参考信号的发送功率；
参考信号使用的天线端口数；
参考信号的频带宽度；
参考信号的频域密度；或，
参考信号的周期。
在该实现方式中，在需要重新收集AI模型的训练数据的情况下，与训练数据收集相关的参考信号的空口传输配置是可以更新的，从而可以改善或保障该参考信号的质量，以便于收集到有效的候选训练数据，从而为AI模型的训练，比如，初始训练/或更新过程的训练，提供保障。另外，由于参考信号的空口传输配置的更新有助于收集到有效的候选训练数据，因此还可以加快AI模型训练的效率。
此外，通过对AI模型的训练数据的收集状况的了解，可以获知是否有足够多的有效候选训练数据进行AI模型的训练/更新，为了保障可靠的基于AI模型的空口性能，有必要对AI模型进行及时的维护，或者切换到非AI的模式，或者自适应作出训练数据收集的配置更新。
结合第一方面，在第一方面的某些实现方式中，所述第三信息还指示所述有效性的判定的最大次数k，k为正整数。
在该实现方式中，通过第三信息指示有效性判定的最大次数k，也即只有在确定需要重新收集的情况下，第二网元才向第一网元指示该有效性判定的最大次数，可以避免在收集结果未知的情况下，就对重新收集的过程进行约束带来的信令浪费。例如，第一网元可能通过一次收集就获得了有效的候选训练数据，此时不需要进行重新收集，此时第二网元就不需要向第一网元指示重新收集的相关信息，以节省信令开销。
结合第一方面，在第一方面的某些实现方式中，所述第一信息还指示所述有效性的判定的最大次数k，k为正整数。
在该实现方式中，通过第一信息指示有效性判定的最大次数k，也即在开始收集训练数据的时候，第二网元就是指示了有效性判定的最大次数，以便于第一网元在一次收集失败之后，可以快速进入重新收集过程，可以节省第一网元和第二网元的交互时间，提高收集训练数据的效率。
在上述两种实现方式中，第二网元通过向第一网元指示有效性的判定的最大次数k，使得第一网元 可以在第一次收集的候选训练数据无效的情况下，快速进行下一次候选训练数据的收集，并且在不超过有效性判定的最大次数k的情况下，可以重复进行候选训练数据的收集，可以节省重新收集的指示信令的开销，同时也提高了AI模型训练/更新的效率。
结合第一方面，在第一方面的某些实现方式中，所述方法还包括：
所述第一网元基于所述更新的空口传输配置，收集所述AI模型的候选训练数据；
若达到所述有效性的判定的最大次数k，且所述第一网元根据所述第一信息确定第k次有效性的判定结果为无效，所述第一网元停止收集所述AI模型的候选训练数据。
在该实现方式中，基于有效性判定的最大次数k的约束，可以避免第一网元的收集过程不陷入死循环，避免资源占用和浪费。
结合第一方面，在第一方面的某些实现方式中，所述方法还包括：
在超过所述有效性的最大判定次数k之前，若所述第一网元根据所述第一信息确定第j次有效性的判定结果为有效，所述第一网元向所述第二网元发送第四信息，所述第四信息包括第二训练数据，且所述第四信息指示所述第j次有效性的判定结果为有效，所述第二训练数据包括所述第j次有效性的判定所针对的候选训练数据中的有效数据，j小于或等于k，j为正整数。
结合第一方面，在第一方面的某些实现方式中，所述第一网元收集所述AI模型的候选训练数据，包括：
所述第一网元测量来自于所述第二网元的参考信号，获得一个或多个测量结果，所述AI模型的候选训练数据包括所述一个或多个测量结果；或者，
所述第一网元测量来自于第三网元的参考信号，获得一个或多个测量结果，所述AI模型的候选训练数据包括所述一个或多个测量结果。
在该实现方式中，基于不同的应用场景，第一网元收集AI模型的候选训练数据，可以是通过测量第二网元发送的参考信号，或者第三网元发送的参考信号，获得的测量结果。可选地，该测量结果可以是一个或多个。示例性地，第一网元通过一次参考信号的测量，获得一个测量结果；或者，第一网元通过多次参考信号的测量，获得多个测量结果；或者，第一网元通过一次参考信号的测量，获得多个测量结果，不作限定。在这些实现中，候选训练数据包括该一个测量结果或该多个测量结果。
结合第一方面，在第一方面的某些实现方式中，第一网元为终端设备，第二网元为接入网设备；所述第一网元测量来自于所述第二网元的参考信号，获得所述一个或多个测量结果。
可选地，来自于第二网元的信号包括如下一项或多项：信道状态信息-参考信号(chanel state information-reference signal,CSI-RS)、定位参考信号(positioning reference signal,PRS)、同步信号和物理广播信道块(synchronizing signal and physical broadcast channel block,SSB)中的同步信号和/或物理广播信道上的信号。
在该实现方式中，AI模型的应用可以适用于基于AI模型的CSI反馈或CSI预测、基于AI模型的波束管理等应用场景，可以解决CSI反馈或预测、波束管理等问题，提高这些应用场景中的空口性能。
结合第一方面，在第一方面的某些实现方式中，所述第一训练数据还包括所述一个或多个测量结果中的K个最优的测量结果对应的参考信号的信息，K为大于或等于1的整数。应理解，当测量结果为1个时，K即等于1；当测量结果为V个，K小于或等于V，且K大于或等于1，其中，V为大于或等于2的整数。
在该实现方式中，AI模型适用于波束管理的场景下，此时，第一训练数据还包括K个最优的测量结果对应的参考信号的信息，用于作为AI模型的标签。
结合第一方面，在第一方面的某些实现方式中，所述第一网元为接入网设备，所述第二网元为定位设备；
所述第一网元测量来自于所述第三网元的探测参考信号，获得所述一个或多个测量结果；
以及，所述第一训练数据还包括所述第三网元的位置信息。
在该实现方式中，AI模型适用于上行定位的场景下，此时，第一网元测量第三网元的探测参考信号，获得候选训练数据，该候选训练数据包括第三网元的位置信息。在候选训练数据有效的情况下，第一网元将有效的候选训练数据(即第一训练数据)和对应的第三网元的位置信息提供给定位设备，以用于AI模型的训练或更新，其中，第三网元的位置信息作为AI模型的标签。
结合第一方面，在第一方面的某些实现方式中，所述第一网元为终端设备，所述第二网元为定位设备；
所述第一网元测量来自于第三网元的定位参考信号，获得所述一个或多个测量结果，所述第三网元为接入网设备；
以及，所述第一训练数据还包括所述第一网元的位置信息。
在该实现方式中，AI模型适用于下行定位的场景下，如果第一网元(例如位置参考设备)收集的候选训练数据有效，第一网元向第二网元(即定位设备)提供的第一训练数据还包括第一网元的位置信息，第一网元的位置信息用于作为AI模型的标签。
第二方面，提供了一种AI模型训练中用于获取训练数据的方法，可以应用于AI模型的训练网元，例如接入网设备或定位设备，该方法包括：
第二网元向第一网元发送第一信息，所述第一信息用于所述第一网元收集的所述AI模型的候选训练数据的有效性的判定，所述有效性的判定结果包括有效或无效；
所述第二网元接收来自于所述第一网元的第二信息，所述第二信息指示所述有效性的判定结果。
其中，针对第一信息的描述，可以参考第一方面中的描述，在此不予赘述。
结合第二方面，在第二方面的某些实现方式中，所述第二信息包括第一训练数据且所述第二信息指示所述第一网元收集的所述候选训练数据有效，所述第一训练数据为所述候选训练数据中的有效数据。
结合第二方面，在第二方面的某些实现方式中，所述第二信息指示所述第一网元收集的所述候选训练数据无效。
结合第二方面，在第二方面的某些实现方式中，所述第一信息用于所述第一网元收集的所述候选训练数据的有效性的判定的约束条件的确定。
结合第二方面，在第二方面的某些实现方式中，若所述候选训练数据中包含满足所述约束条件的第一训练数据，所述候选训练数据有效；或者，
若所述候选训练数据中不包含满足所述约束条件的第一训练数据，所述候选训练数据无效。
结合第二方面，在第二方面的某些实现方式中，在所述第二信息指示所述第一网元收集的所述候选训练数据无效的情况下，所述方法还包括：
所述第二网元向所述第一网元发送第三信息，所述第三信息指示所述第一网元重新收集所述AI模型的候选训练数据。
结合第二方面，在第二方面的某些实现方式中，所述方法还包括：
所述第二网元确定空口传输配置信息，所述空口传输配置信息对应更新的空口传输配置，所述空口传输配置信息指示所述第一网元基于所述更新的空口传输配置收集所述AI模型的候选训练数据；
其中，所述更新的空口传输配置信息包括如下一项或多项的更新：
参考信号的发送功率；
参考信号使用的天线端口数；
参考信号的频带宽度；
参考信号的频域密度；或，
参考信号的周期。
结合第二方面，在第二方面的某些实现方式中，所述第三信息还指示所述有效性的判定的最大次数k，k为正整数。
结合第二方面，在第二方面的某些实现方式中，所述第一信息还指示所述有效性的判定的最大次数k，k为正整数。
结合第二方面，在第二方面的某些实现方式中，所述方法还包括：
所述第二网元接收来自于所述第一网元的第四信息，所述第四信息包括第二训练数据，且所述第四信息指示所述第一网元的第j次有效性判定的判定结果为有效，所述第二训练数据为所述第j次有效性的判定所针对的候选训练数据中的有效数据，j小于或等于k，j为正整数。
结合第二方面，在第二方面的某些实现方式中，所述第二网元为接入网设备，所述第一网元为终端设备，所述方法还包括：
所述第二网元向所述第一网元发送参考信号，所述参考信号用于所述第一网元获取对应于所述参考信号的一个或多个测量结果，所述AI模型的候选训练数据包括所述一个或多个测量结果。
结合第二方面，在第二方面的某些实现方式中，所述第一训练数据还包括所述一个或多个测量结果中的K个最优的测量结果对应的参考信号，K为大于或等于1的整数。
结合第二方面，在第二方面的某些实现方式中，所述第二网元为定位设备，所述第一网元为接入网设备，所述AI模型的候选训练数据包括一个或多个测量结果和第三网元的位置信息，所述一个或多个测量结果是由所述第一网元测量所述第三网元发送的探测参考信号获得的。
结合第二方面，在第二方面的某些实现方式中，所述第二网元为定位设备，所述第一网元为终端设备，所述AI模型的候选训练数据包括一个或多个测量结果和所述第一网元的位置信息，所述一个或多个测量结果基于对所述第三网元发送的定位参考信号的测量，所述第三网元为接入网设备。可选的，该测量由第一网元执行。
结合第二方面，在第二方面的某些实现方式中，所述约束条件包括如下一项或多项：
质量指标的门限和所述质量指标的判定准则；或，
符合质量指标的判定准则的训练数据的数量门限和所述训练数据的数量的判定准则；或，
单次有效性判定对应的候选训练数据收集的最大时长。
结合第二方面，在第二方面的某些实现方式中，所述第一信息指示如下一项或多项：
质量指标的门限；
质量指标的判定准则；
符合质量指标的判定准则的训练数据的数量门限；
符合质量指标的判定准则的训练数据的数量的判定准则；或
单次有效性判定对应的候选训练数据收集的最大时长。
可选地，上述实现方式中的质量指标包括一项或多项质量指标，比如，包括AI模型的标签的质量指标，或，对参考信号的测量结果的质量指标等中的一项或多项。
在第一方面或第二方面的某些实现方式中，所述约束条件基于所述AI模型的应用场景，所述AI模型的应用场景包括如下一项或多项：
基于所述AI模型的CSI反馈或CSI预测、基于所述AI模型的定位，或，基于所述AI模型的波束管理。
第三方面，本申请提供一种通信装置，该通信装置可以是终端设备，也可以是设置于终端设备中的装置、模块或芯片等，或者是能够和终端设备匹配使用的装置。一种设计中，该通信装置可以包括用于执行第一方面所述的方法/操作/步骤/动作所一一对应的模块，该模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。一种设计中，该通信装置可以包括处理模块和通信模块。
第四方面，本申请提供一种通信装置，一种设计中，该通信装置可以包括用于执行第二方面所述的方法/操作/步骤/动作所一一对应的模块，该模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。一种设计中，该通信装置可以包括处理模块和通信模块。在一个示例中，该通信装置为接入网设备或定位设备，定位设备例如可以为LMF网元。
第五方面，本申请提供一种通信装置，所述通信装置包括处理器，用于实现上述第一方面或第一方面的任一实现方式中所述的方法。处理器与存储器耦合，存储器用于存储指令和数据，所述处理器执行所述存储器中存储的指令时，可以实现上述第一方面或第一方面的任一实现方式中所述的方法。可选的，所述通信装置还可以包括存储器。可选的，所述通信装置还可以包括通信接口，所述通信接口用于该装置与其它设备进行通信，示例性的，通信接口可以是收发器、硬件电路、总线、模块、管脚或其它类型的通信接口。在一个示例中，该通信装置可以是终端设备，也可以是用于设置于终端设备中的装置、模块或芯片等，或者是能够和终端设备匹配使用的装置。
第六方面，本申请提供一种通信装置，所述通信装置包括处理器，用于实现上述第二方面或第二方面的任一实现方式中所述的方法。处理器与存储器耦合，存储器用于存储指令和数据，所述处理器执行所述存储器中存储的指令时，可以实现上述第二方面或第二方面的任一实现方式中所述的方法。可选的，所述通信装置还可以包括存储器。可选的，所述通信装置还可以包括通信接口，所述通信接口用于该装置与其它设备进行通信，示例性的，通信接口可以是收发器、硬件电路、总线、模块、管 脚或其它类型的通信接口。在一个示例中，该通信装置可以为接入网设备，也可以是用于设置于接入网设备中的装置、模块或芯片等，或者是能够和接入网设备匹配使用的装置。在另一个示例中，该通信装置可以为定位设备，也可以是用于设置于定位设备中的装置、模块或芯片等，或者是能够和定位设备匹配使用的装置。
第七方面，本申请提供一种通信系统，包括第一网元和第二网元。示例性地，第一网元和第二网元之间的交互如下：
第二网元向第一网元发送第一信息，所述第一信息用于所述第一网元收集的候选训练数据的有效性的判定，所述有效性的判定结果包括有效或无效；
所述第一网元接收来自于所述第二网元的所述第一信息；
所述第一网元收集AI模型的候选训练数据；
所述第一网元根据所述候选训练数据和所述第一信息，向所述第二网元发送第二信息，所述第二信息指示有效性的判定结果；
所述第二网元接收来自于所述第一网元的所述第二信息。
具体地，第一网元侧的方案可以参考第一方面中实现进行理解，第二网元侧的方案可以参考第二方面的实现进行理解，这里不再赘述。示例性地，该通信系统包括终端设备和接入网设备。可选地，该通信系统包括终端设备、接入网设备和定位设备。可选地，终端设备为位置参考设备，定位设备为LMF网元。
第八方面，本申请提供一种通信系统，包括如第三方面或第五方面所述的通信装置，以及如第四方面或第六方面所述的通信装置。
第九方面，本申请还提供了一种计算机程序，当所述计算机程序在计算机上运行时，使得所述计算机执行上述第一方面、第二方面，或者，如第一方面或第二方面的任一实现方式中提供的方法。
第十方面，本申请还提供了一种计算机程序产品，包括指令，当所述指令在计算机上运行时，使得计算机执行上述第一方面、第二方面，或者，如第一方面或第二方面的任一实现方式中提供的方法。
第十一方面，本申请还提供了一种计算机可读存储介质，所述计算机可读存储介质中存储有计算机程序或指令，当所述计算机程序或者指令在计算机上运行时，使得所述计算机执行上述第一方面、第二方面，或者，如第一方面或第二方面的任一实现方式中提供的方法。
第十二方面，本申请还提供了一种芯片，所述芯片用于读取存储器中存储的计算机程序，执行上述第一方面、第二方面，或者，如第一方面或第二方面的任一方面提供的方法；或者，所述芯片包括用于执行上述第一方面、第二方面，或者，如第一方面或第二方面的任一方面提供的方法的电路。
第十三方面，本申请还提供了一种芯片系统，该芯片系统包括处理器，用于支持装置实现上述第一方面、第二方面，或者如所述第一方面或第二方面中任一方面提供的方法。在一种可能的设计中，所述芯片系统还包括存储器，所述存储器用于保存该装置必要的程序和数据。该芯片系统可以由芯片构成，也可以包含芯片和其他分立器件。
如上第二方面至第十三方面的任一方面或其任一实现方式所提供的方案的技术效果，可参考第一方面中的相应描述，不再赘述。
附图说明
图1为神经网络迭代过程示意图。
图2为适用于本申请实施例的通信系统的架构示意图。
图3为本申请提供的AI模型训练中用于获取训练数据的方法的示意性流程图。
图4为一种基于AI模型的CSI反馈机制的示意图。
图5为本申请提供的基于AI模型的CSI反馈中获取训练数据的示例。
图6为本申请提供的技术方案在基于AI模型的上行定位场景的示意图。
图7为本申请提供的基于AI模型的上行定位中获取训练数据的示例。
图8为本申请提供的技术方案在基于AI模型的下行定位场景的示意图。
图9为本申请提供的基于AI模型的下行定位中获取训练数据的示例。
图10为基于AI辅助的稀疏波束扫描过程示意图。
图11为本申请提供的基于AI模型的波束管理中获取训练数据的示例。
图12为本申请提供的通信装置的示意性结构图。
图13为本申请提供的通信装置的示意性结构图。
具体实施方式
下面将结合附图，对本申请中的技术方案进行描述。
首先对本申请实施例中涉及到的相关概念和技术作简单介绍。
AI模型：指将一定维度的输入映射到一定维度的输出的函数模型，其模型参数通过机器学习训练得到。例如，f(x)＝ax2+b是一个二次函数模型，它可以视作一个AI模型，a和b为该AI模型的参数，可以通过机器学习训练得到。示例性地，本申请下文实施例中提及的AI模型不限于为神经网络、线性回归模型、决策树模型、支持向量机(support vector machine,SVM)、贝叶斯网络、Q学习模型或者其他机器学习(machine learning,ML)模型。
训练数据集：机器学习中用于模型训练、验证和测试的数据，数据的数量和质量将影响到机器学习的效果。训练数据即可以包括AI模型的输入，或者包括AI模型的输入和目标输出。其中，目标输出即为AI模型的输出的目标值，也可以称为输出真值、输出真值、标签或者标签样本。
模型训练：通过选择合适的损失函数，利用优化算法对模型参数进行训练，使得损失函数的取值小于门限，或者使得损失函数的取值满足目标需求的过程。
AI模型设计：主要包括数据收集环节(例如，收集训练数据和/或推理数据)、模型训练环节以及模型推理环节。进一步地还可以包括推理结果应用环节。在前述数据收集环节中，数据源(data source)用于提供训练数据集和推理数据。在模型训练环节中，通过对数据源提供的训练数据(training data)进行分析或训练，得到AI模型。其中，AI模型表征了模型的输入和输出之间的映射关系。通过模型训练节点学习得到AI模型，相当于利用训练数据学习得到模型的输入和输出之间的映射关系。在模型推理环节中，使用经由模型训练环节训练后的AI模型，基于数据源提供的推理数据进行推理，得到推理结果。该环节还可以理解为：将推理数据输入到AI模型，通过AI模型得到输出，该输出即为推理结果。该推理结果可以指示：由执行对象使用(执行)的配置参数、和/或由执行对象执行的操作。在推理结果应用环节中进行推理结果的发布，例如推理结果可以由执行(actor)实体统一规划，例如执行实体可以发送推理结果给一个或多个执行对象(例如，核心网设备、接入网设备、或终端设备等)去执行。又如执行实体还可以反馈模型的性能给数据源，便于后续实施模型的更新训练。
损失函数：用于衡量模型的预测值和真实值之间的差异或差距。
模型应用：利用训练好的模型去解决实际问题。
机器学习(machine learning,ML)是实现人工智能(artificial intelligence,AI)的一种重要技术途径。机器学习可以分为监督学习、非监督学习、强化学习。
作为一个示例，监督学习依据已采集到的样本值和样本标签，利用机器学习算法学习样本值到样本标签的映射关系，并用机器学习模型来表达学到的映射关系。训练机器学习模型的过程就是学习这种映射关系的过程。例如信号检测中，含噪声的接收信号即为样本，该信号对应的真实星座点即为标签。机器学习期望通过训练学习到样本与标签之间的映射关系，即，使机器学习模型学到一种信号检测器。在训练时，通过计算模型的预测值与真实标签的误差来优化模型参数。一旦映射关系学习完成，就可以利用学到的映射关系来预测每一个新样本的样本标签。监督学习学到的映射关系可以包括线性映射、非线性映射。根据标签的类型可将学习的任务分为分类任务和回归任务。
参见图1，图1为神经网络迭代过程示意图。如图1所示，选择n个样本组成一个batch，然后将batch丢进神经网络，得到输出结果。再将输出结果与样本标签丢给loss函数，计算出本轮的loss。最后将每个参数的导数配合步长参数来进行参数更新。这就是训练过程的一次迭代。batch是“批”的意思，即是说神经网络处理数据是分批处理的。batch size就是每批处理的样本的个数。所以一般找一个合适大小的样本量，可以并行计算加快训练速度，而一次处理的数据量又不会过大。
训练数据集是训练样本的集合，每个训练样本为神经网络的一次输入，训练数据集用于模型训练。训练数据集是机器学习最重要的部分之一，机器学习的训练过程本质上就是从训练数据集中学习它的某些特征，从而使得在该训练数据集下，神经网络的输出与理想目标值(也即标签或输出真值)之间 的差异最小。通常情况下，即使采用相同的网络结构，使用不同训练数据集训练出来的神经网络的权重以及输出都不相同。因此，训练数据集的构成与选取，在某种程度上决定了训练出来的神经网络的性能。
当AI模型应用于空口技术中，无论是离线的模型更新/训练，还是在线的模型更新/训练，都需要对真实部署网络中的数据进行收集，以构成模型更新/训练所需的训练数据集。优良的训练数据集有助于无线通信AI算法设计获得更大的性能增益，且提升最终设计算法在多种场景下的泛化能力和鲁棒性。
在AI模型应用于空口技术的一些应用场景下时，若AI模型的训练网元和训练数据的收集网元不在一个网元，AI模型训练网元和训练数据收集网元之间需要进行训练数据的交互。基于当前的技术现状，训练数据的交互通常是周期或持续存在的，容易造成空口资源的浪费。
另外，训练数据收集网元收集训练数据的过程，也未受到AI模型训练网元的需求的约束，经常出现无效收集的情况出现，例如，训练数据收集网元所收集的训练数据不是AI模型训练网元所真正需求的训练数据，导致一些无效交互，带来空口资源的浪费。此外，如果AI模型训练网元使用这些训练数据进行AI模型训练，容易对AI模型的训练数据集造成污染，导致增益评估不准、出现模型过拟合、泛化能力弱、场景适应性差等多种问题。
针对上述问题，本申请提供一种AI模型训练中用于获取训练数据的方法，有益于上述问题的解决或改善。
本申请提供的技术方案可以应用于各种通信系统，例如，该通信系统可以是第四代(4th generation，4G)通信系统(例如长期演进(long term evolution，LTE)系统)、第五代(5th generation，5G)通信系统、全球互联微波接入(worldwide interoperability for microwave access，WiMAX)或者无线局域网(wireless local area network，WLAN)系统、卫星通信系统，或者是未来的通信系统，例如6G通信系统，或者多种系统的融合系统等。其中，5G通信系统还可以称为新无线(new radio,NR)系统。
通信系统中的一个网元可以向另一个网元发送信号，或者从另一个网元接收信号。其中信号可以包括信息、信令或者数据等。其中，网元也可以被替换为实体、网络实体、设备、通信设备、通信模块、节点、通信节点等等，本申请中以网元为例进行描述。
适用于本申请的通信系统，可以包括第一网元和第二网元，可选地，还包括第三网元。其中，关于第一网元、第二网元以及第三网元的数量不作限定。
参见图2，图2为适用于本申请实施例的通信系统的架构示意图。图2的(a)为适用于本申请实施例的一种通信系统的架构示意图。示例性地，该通信系统中包括网络设备110，终端设备120和终端设备130。终端设备120和130可以接入网络设备110，并和网络设备110进行通信。可选地，网络设备110可以为接入网设备。在一种实现中，该通信系统中还可以包括AI实体，网络设备可已将终端设备上报的与AI模型相关的数据转发给AI实体，由AI实体执行训练数据集构建、模型训练等AI相关的操作，并将训练好的AI模型、模型评估、测试结果等AI相关操作的输出提供给网络设备。在另一种实现中，AI实体也可以位于网络设备110内部，即为网络设备110的一个模块。图2的(b)为适用于本申请实施例的另一种通信系统的架构示意图。该通信系统中包括网络设备110，终端设备120，终端设备130和定位设备140。其中，定位设备140和网络设备110之间可以通过接口消息进行通信。示例性地，定位设备140为定位管理功能(location management function,LMF)，网络设备110可以为接入网设备，例如gNB或eNB等，不作限定。示例性地，若接入网设备110为gNB，则gNB和LMF之间可以通过NR定位协议A(NR positioning protocol A,NRPPa)消息交互信息；若接入网设备110为eNB，则eNB和LMF之间可以通过LTE定位协议(LTE positioning protocol,LPP)消息交互信息。可选地，终端设备与定位设备140之间也可以直接进行通信，如图2的(b)中所示的终端设备130与定位设备140之间的交互。在图2中，AI实体可以配置在定位设备140内部，或者和定位设备140分离配置，不作限定。可选的，定位设备可以与网络设备为同一个设备的不同模块，也可以是分离的不同设备。
在实际应用中，一个网络设备可以同时服务于一个或多个终端设备。一个终端设备也可以同时接入一个或多个网络设备。本申请实施例对该无线通信系统中包括的终端设备和网络设备的数量不做限定。此外，对于图2的(b)的定位设备140，也不限于为LMF网元，还可以为其它具有定位功能的网元，对其数量也不作限定。
示例性地，网络设备可以是具有无线收发功能的设备，该网络设备可以是提供无线通信功能服务的设备，通常位于网络侧，包括但不限于第五代(5th generation，5G)通信系统中的下一代基站(gNodeB,gNB)、第六代(6th generation,6G)移动通信系统中的基站、未来移动通信系统中的基站，或无线保真(wireless fidelity,WiFi)系统中的接入节点(access point,AP)，长期演进(long term evolution,LTE)系统中的演进型节点B(evolved node B,eNB)、无线网络控制器(radio network controller,RNC)、节点B(node B,NB)、基站控制器(base station controller,BSC)、家庭基站(例如,home evolved NodeB,或home Node B,HNB)、基带单元(base band unit,BBU)，传输接收点(transmission reception point,TRP)、发射点(transmitting point,TP)、基站收发台(base transceiver station,BTS)、卫星、无人机等。在一种网络结构中，网络设备可以包括集中单元(centralized unit,CU)节点，或包括分布单元(distributed unit,DU)节点，或者为包括CU节点和DU节点的RAN设备，或者为包括控制面CU节点和用户面CU节点，以及DU节点的RAN设备，或者，网络设备还可以为云无线接入网络(cloud radio access network,CRAN)场景下的无线控制器、中继站、车载设备以及可穿戴设备等。此外，基站可以是宏基站、微基站、中继节点、施主节点，或其组合。基站还可以指用于设置于前述设备或装置内的通信模块、调制解调器或芯片。基站还可以是移动交换中心以及设备到设备(device to device,D2D)、车辆外联(vehicle-to-everything,V2X)、机器到机器(machine to machine,M2M)通信中承担基站功能的设备、6G网络中的网络侧设备、未来的通信系统中承担基站功能的设备等。基站可以支持相同或不同接入技术的网络，不作限定。
网络设备可以是固定的，也可以是移动的。例如，接入网设备110可以是静止的，并负责来自终端设备120和130的一个或多个小区中的无线传输和接收。接入网设备110也可以是移动的，例如，直升机或无人机可以被配置成充当移动基站，并且一个或多个小区可以根据移动基站的位置移动。应理解，在其它示例中，直升机或无人机可以被配置成用作与基站110通信的设备。
本申请中，用于实现如上接入网络功能的通信装置可以是接入网设备，也可以是具有接入网络的部分功能的网络设备，也可以是能够支持实现接入网络功能的装置，例如芯片系统，硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在接入网设备中或者和接入网设备匹配使用。本申请的方法中，以用于实现接入网设备功能的通信装置是接入网设备为例进行描述。
终端设备可以是用户侧的一种用于接收或发射信号的实体，如手机。终端设备包括具有无线连接功能的手持式设备、连接到无线调制解调器的其他处理设备或车载设备等。终端设备可以是便携式、袖珍式、手持式、计算机内置的或者车载的移动装置。终端设备120可以广泛应用于各种场景，例如蜂窝通信、WiFi系统、D2D、V2X、端到端(peer to peer,P2P)、M2M、机器类型通信(machine type communication,MTC)、物联网(internet of things,IoT)、虚拟现实(virtual reality,VR)、增强现实(augmented reality,AR)、工业控制、自动驾驶、远程医疗、智能电网、智能家具、智能办公、智能穿戴、智能交通、智慧城市、无人机、机器人、遥感、被动传感、定位、导航与跟踪、自主交付与移动等。通信设备120的一些举例为：3GPP标准的用户设备(user equipment,UE)、WiFi系统中的站点(station,STA)、固定设备、移动设备、手持设备、可穿戴设备、蜂窝电话、智能电话、会话发起协议(session initialization protocol,SIP)电话、笔记本电脑、个人计算机、智能书、车辆、卫星、全球定位系统(global positioning system,GPS)设备、目标跟踪设备、无人机、直升机、飞行器、船只、遥控设备、智能家居设备、工业设备、个人通信业务(personal communication service，PCS)电话、无线本地环路(wireless local loop，WLL)站、个人数字助理(personal digital assistant，PDA)、无线网络摄像头、平板电脑、掌上电脑、移动互联网设备(mobile internet device，MID)、可穿戴设备如智能手表、虚拟现实(virtual reality，VR)设备、增强现实(augmented reality，AR)设备、工业控制(industrial control)中的无线终端、车联网系统中的终端、无人驾驶(self driving)中的无线终端、智能电网(smart grid)中的无线终端、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端如智能加油器，高铁上的终端设备以及智慧家庭(smart home)中的无线终端，如智能音响、智能咖啡机、智能打印机等。终端设备120可以为以上各种场景中的无线设备或用于设置于无线设备的装置，例如，上述设备中的通信模块、调制解调器或芯片等。终端设备也可以称为终端、用户设备(user equipment,UE)、移动台(mobile station，MS)、移动终端(mobile terminal，MT)等。终端设备还可以是未来的无线通信系统中的终端设备。此外，终端设备还可以包括位置参考设备，例如，自动导航 小车(automated guided vehicle,AGV)或具有类似功能的设备。本申请的实施例对终端设备所采用的具体技术和具体设备形态不做限定。
本申请中，用于实现终端设备功能的通信装置可以是终端设备，也可以是具有以上通信设备的部分功能的终端设备，也可以是能够支持实现以上终端设备的功能的装置，例如芯片系统，该装置可以被安装在终端设备中或者和终端设备匹配使用。本申请中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。
应理解，图2所示的通信系统中各个设备的数量、类型仅作为示意，本申请并不限于此，实际应用中在通信系统中还可以包括更多的终端设备、更多的接入网设备、更多的定位设备，还可以包括其它网元，例如可以包括核心网设备，和/或用于实现人工智能功能的网元。
下面介绍本申请提供的技术方案。
参见图3，图3为本申请提供的AI模型训练中用于获取训练数据的方法的示意性流程图。在图3的方法中，第一网元可以为收集AI模型的训练数据的网元，第二网元为AI模型的训练网元。可选地，第二网元可以是AI模型的训练网元，同时还是AI推理发生的网元。第一网元和第二网元可以是逻辑上分离部署的，作为不同的实现方式，第一网元和第二网元可能在物理上部署于同一个网元或不同的网元，不作限定。
310、第一网元接收来自于第二网元的第一信息，第一信息指示用于第一网元收集的候选训练数据的有效性的判定，其中，有效性的判定结果包括有效或无效。
第一网元根据第一信息，可以对收集的候选训练数据的有效性进行判定，也或者说，根据第一信息，第一网元可以判断所收集的候选训练数据是否有效。
具体地，第一网元在根据第一信息判定所收集的候选训练数据包含有效的训练数据的情况下，第一网元则确定所收集的候选训练数据有效；若第一网元根据第一信息判定所收集的候选训练数据中不包含有效的候选训练数据，第一网元则判定所收集的候选训练数据无效。换句话说，若第一网元所收集的候选训练数据中部分候选训练数据有效，即判定结果为有效。在判定结果为有效的情况下，有效的候选训练数据即成为第一网元此次收集的训练数据。示例性地，有效的候选训练数据(以下可简称为“有效数据”)可以是所收集的候选训练数据中的部分或全部，不作限定，在本文中统称为第一训练数据。若第一网元所收集的候选训练数据中不包含有效的候选训练数据，则判定结果为无效。
可以发现，在本申请实施例中，第一网元在收集到候选训练数据之后，会在第一网元侧进行一次有效性的判定。
作为一个示例，第一信息指示约束条件，该约束条件用于第一网元判定所收集的候选训练数据的有效性。
示例性地，约束条件包括如下一项或多项：
质量指标的门限和所述质量指标的判定准则；或，
符合质量指标的判定准则的训练数据的数量门限和所述训练数据的数量的判定准则，或，
单次有效性判定对应的候选训练数据收集的最大时长。
可选地，第一信息用于约束条件的确定。
示例性地，第一信息指示如下一项或多项信息：
质量指标的门限；
质量指标的判定准则；
符合质量指标的判定准则的训练数据的数量的门限；
符合质量指标的判定准则的训练数据的数量的判定准则；或者，
单次有效性判定对应的候选训练数据收集的最大时长。
假设约束条件包括一个或多个质量指标，第一信息用于约束条件的确定时，可以有多种实现方式。
可选地，在一个示例中，第一信息指示该一个或多个质量指标的门限，该一个或多个质量指标的判定准则是由协议预定义的。在该示例中，第一网元根据第一信息和协议预定义，确定该约束条件。
可选地，在另一个示例中，第一信息指示该一个或多个质量指标的门限，以及该一个或多个质量指标的判定准则。在该示例中，第一网元根据第一信息，确定该约束条件。
可选地，在再一个示例中，约束条件包括多个质量指标，第一信息指示该多个质量指标中部分质 量指标的门限，另一部分质量指标的门限以及该多个质量指标的判定准则由协议预定义。在该示例中，第一网元根据第一信息和协议预定义，确定约束条件。
可选地，在再一个示例中，第一信息指示部分质量指标的门限和一个索引信息，该索引信息用于确定所述部分质量指标的判定准则，以及约束条件中其它质量指标的门限以及所述其它质量指标的判定准则。在该示例中，第一网元根据第一信息和该索引信息，确定约束条件。
可选地，在再一个示例中，第一信息指示一个索引信息，该索引信息用于确定一个或多个质量指标的门限，以及所述一个或多个质量指标的判定准则。在该示例中，第一网元根据该索引信息，确定约束条件。
此外可选地，上述示例中的“协议预定义”也可以为预先配置或预先存储等其它实现方式，不作限定。
此外，关于第一信息用于确定约束条件的具体实现还可以参考发明内容部分的相应实现，这里不再重复说明。
在一个示例中，第一信息还指示有效性判定的最大次数k，k为正整数。
可以理解的是，第一信息包括如上信息中的多项时，该多项信息可以携带在一条消息或分别携带在多条消息中，也即，第一信息可以携带在一条消息或携带在多条消息中。
其中，单次有效性判定对应的候选训练数据收集的最大时长，在下文记作Z，Z为大于0的数。该候选训练数据收集的最大时长，也即，候选训练数据可用于有效性判定的最大时长，候选训练数据的保留时长超过该最大时长Z，候选训练数据失效，不再用于有效性判定。可选的，该最大时长Z可以与两次相邻的有效性判定的间隔时间相同，或者，可以大于或小于两次相邻的有效性判定的间隔时间。在该最大时长大于两次相邻的有效性判定的间隔时间的情况下，一次有效性判定对应的候选训练数据可以包括该次有效性判定的前一次或多次有效性判定所对应的候选训练数据中的全部或部分。所述两次相邻的有效性判定的间隔时间可以是固定的，也即，有效性判定在一定时间内周期进行，也可以是变化的，也即，有效性判定时间不固定，比如，有效性判定是对满足门限要求的候选训练数据进行计数，当满足门限要求的候选训练数据的数量满足要求时，该次有效性判定完成，判定结果为有效；当满足门限要求的候选训练数据的数量不满足要求且超过了有效性判定的最大间隔时间T(即预设的间隔时间的门限)或者所收集的候选训练数据的数量超过预设门限(即所收集的候选训练数据的最大数量)，该次有效性判定也完成，判定结果为无效。具体的有效性判定的时间相关的信息，比如判定时间，周期性判定的起始时间，或，周期，最大间隔时间T，所收集的候选训练数据的最大数量等中的一项或多项，可以全部或部分由协议预定义，或者，基于配置。
根据上文的说明已知，在本申请中，第二网元向第一网元发送第一信息，第一信息用于第一网元收集的候选训练数据的有效性判定，实际上，第二网元通过第一信息向第二网元指示了训练数据的要求。或者说，符合该要求的候选训练数据才能作为训练数据用于AI模型的训练或更新，可见，第一网元所收集的候选训练数据需要经过“筛选”之后，符合要求的候选训练数据才会作为训练数据，由第一网元提供给第二网元使用。因此，第一网元收集到候选训练数据之后，会根据第一信息对收集到的候选训练数据进行有效性的判定。如果判定结果为无效，代表此次收集的候选训练数据不是第二网元所需求的，也即此次收集的候选训练数据中不包含符合要求的候选训练数据。在此情况下，第一网元可能会涉及到训练数据的重新收集。因此，第一网元在收集AI模型的训练数据的过程中，训练数据可能并不是一次收集就能获得的。在具体的实现中，单次有效性判定发生的最大时间间隔T相当于指定了第一网元多久执行一次有效性判定。
举例来说，若第i次有效性判定的判定结果为无效，第一网元重新收集候选训练数据。经过一段时间间隔，第一网元对所收集的候选训练数据进行第i+1次有效性的判定，i为正整数。因此，在本申请实施例中，第一网元收集AI模型的候选训练数据的次数和第一网元执行判定有效性判定的次数是对应的，或者说，是相等的。换句话说，第一网元每执行一次有效性的判定，代表着在此次判定之前有一次候选训练数据的收集。为了技术方案描述上的清楚，文中将第i个判定之前的一次候选训练数据的收集称为第i次收集。考虑到不同的实现方式，在该示例中，第i+1次的有效性判定可以是针对第i+1次收集获得的候选训练数据的，也可以是针对第i+1次收集以及第i+1次收集之前的一次或多次收集获得的候选训练数据的，不作限定。在此实现下，可能涉及到在一次有效性判定之后，第一网元如何处理 该次判定之前所收集的候选训练数据的问题。
作为一个示例，第一网元开始候选训练数据的第i次收集之后，经过时间间隔T0(小于或等于最大时间间隔T)，对第i次收集的候选训练数据进行有效性判定，也即第i次有效性判定。假设第i次有效性判定的判定结果为无效，第一网元可以丢弃第i次收集的候选训练数据，在未超过有效性判定的最大次数k的情况下，再进行第i+1次收集。在该示例中，每一次有效性判定只针对时间间隔T0内收集到的候选训练数据，当该次收集无效，则丢弃该次收集的候选训练数据。作为另一个示例，一次有效性判定可以针对多个时间间隔T0内收集到的候选训练数据。或者说，一个有效性判定所针对的候选训练数据可能来自于多次收集。假设第i次有效性判定的判定结果为无效，在未超过有效性判定的最大次数k的情况下，第一网元可以保留第i次收集的部分候选训练数据。例如，第一网元保留第i次收集的候选训练数据中满足约束条件中的部分质量指标的判定准则的这部分候选训练数据。之后，再进行第i+1次收集。经过时间间隔T0，第一网元针对第i+1次收集的候选训练数据以及所保留时长未超过最大时长Z的第i次收集的部分候选训练数据进行有效性判定，即第i+1次有效性判定。可选地，在未超过有效性判定的最大次数k之前，若有效性的判定结果一直为无效，则第一网元可以将每次采集的候选训练数据中符合部分质量指标的判定准则的候选训练数据保留下来，并在完成一次新的收集之后，将保留的符合部分质量指标的判定准则的历史候选训练数据和新收集的候选训练数据一起进行有效性的判定。由此可见，在一次有效性判定中，被确定为无效的候选训练数据是针对该次有效性判定而言的，不代表此次有效性判定中被确定为无效的候选训练数据永远不能作为训练数据。本申请对这些具体的实现方式不作限定。
作为一个示例，约束条件基于AI模型的应用场景。例如，AI模型的应用场景包括但不限于为如下场景：
基于AI模型的CSI反馈或CSI预测、基于AI模型的定位，或者基于AI模型的波束管理。
在不同的应用场景下，约束条件中涉及的质量指标、质量指标的门限、质量指标的判定准则、符合质量指标的判定准则的训练数据的数量的门限，以及所述训练数据的数量的判定准则中的一项或多项可以不同，下文会针对不同的应用场景分别作举例说明。
320、第一网元收集AI模型的候选训练数据。
第一网元收集AI模型的候选训练数据。可选地，第一网元可以在接收到第一信息之后，也即，基于第一信息的触发，开始收集AI模型的候选训练数据。可选的，第一网元也可以在接收第一信息之前或同时，开始收集该AI模型的候选训练数据。也即，步骤310和步骤320的发生先后可以不予限定。
可选地，考虑到第二网元所需求的训练数据的要求可能是变化的，当第二网元对训练数据的要求发生变化，第二网元可以向第一网元发送更新的第一信息。其中，第一信息的更新主要是指根据第一信息所确定的约束条件的更新。当第一网元接收到更新的第一信息，则基于更新的第一信息所确定的约束条件，对收集到的候选训练数据进行有效性的判定。下文各实施例中仅以第一网元某一次接收到的第一信息作为示例，对有效性判定以及后续的流程进行说明。
330、第一网元向第二网元发送第二信息，第二信息指示第一网元收集的候选训练数据的有效性的判定结果。
若第一网元根据第一信息确定所收集的候选训练数据有效，第一网元向第二网元发送第二信息，第二信息指示第一网元收集的候选训练数据有效。
可选地，在一个示例中，第一网元向第二网元发送第一训练数据，第一训练数据本身隐含指示第一网元收集的候选训练数据有效。可选地，在另一个示例中，第一网元向第二网元发送第一训练数据，此时，第一网元还向第二网元发送用于指示第一网元收集的候选训练数据有效的信息，例如，信息a。在该示例中，第一网元向第二网元发送第一训练数据和信息a，信息a指示此次收集有效。
若第一网元根据第一信息确定所收集的候选训练数据无效，第一网元向第二网元发送第二信息，第二信息指示第一网元收集的候选训练数据无效。应理解，在第一网元收集的候选训练数据无效的情况下，第一网元仅向第二网元发送所收集的候选训练数据无效的指示，而不发送所收集的无效的候选训练数据，由此可以降低空口资源的浪费。
作为一个示例，在第一网元确定收集的候选训练数据无效的情况下，第一网元丢弃本次所收集的候选训练数据；或者，在上文描述的一些实现中，一次有效性判定中，无效的候选训练数据也可以被 保留，用于后续的有效性判定。进一步地，若第二网元指示第一网元重新收集AI模型的训练数据，第一网元重新收集AI模型的候选训练数据。
可选地，在步骤320中，第一网元收集AI模型的候选训练数据，具体可以是第一网元测量来自于第二网元或第三网元的参考信号，获得AI模型的候选训练数据，也或者说，该候选训练数据包括第一网元测量参考信号获得的测量结果。在本申请中，参考信号泛指用于信道测量的信号。该信道测量可以用于信道状态信息反馈，波束管理，或定位等功能中的一项或多项。所述参考信号可以包括信道状态信息参考信号，同步信号，如主同步信号和/或辅同步信号，物理广播信号，同步信号和物理广播信号块(SSB)，解调参考信号，相位跟踪参考信号，或，定位参考信号中的一项或多项。当AI模型应用于不同的场景中时，参考信号可能是不同的，下文的实施例会针对不同的应用场景分别举例说明。此外，上述第三网元是指不同于第二网元的一个网元。
可以理解的是，在一个示例中，第一网元测量来自于第二网元的参考信号，获得测量结果。可选地，该测量结果可以是一个或多个。第一网元收集的AI模型的候选训练数据包括该一个或多个测量结果。在该示例中，第二网元在向第一网元发送参考信号之前，第二网元向第一网元发送空口传输配置信息，该空口传输配置信息对应空口传输配置，该空口传输配置信息指示第一网元基于该空口传输配置收集AI模型的候选训练数据。换句话说，第二网元是根据该空口传输配置发送参考信号的，第一网元测量来自于第二网元的参考信号，获得测量结果，从而收集到基于该空口传输配置的候选训练数据。可选地，第一网元为UE，第二网元为接入网设备，例如基站。
在另一个示例中，第一网元测量来自于第三网元的信号，获得测量结果。可选地，该测量结果可以是一个或多个。第一网元收集的AI模型的候选训练数据包括该一个或多个测量结果。可选地，第一网元为接入网设备，例如基站，第三网元为UE。在该示例中，第一网元在测量来自于第三网元的参考信号之前，第一网元配置第三网元发送参考信号，具体地，第一网元向第三网元发送空口传输配置信息，该空口传输配置信息对应空口传输配置。与上一个示例类似，第三网元基于该空口传输配置发送参考信号，第一网元测量来自于第三网元的参考信号，获得测量结果，从而获得基于该空口传输配置的候选训练数据。
在上述示例中，可选地，该空口传输配置可以包括如下一项或多项：
参考信号的发送功率；
参考信号所使用的天线端口数；
参考信号的频带宽度；
参考信号的频域密度；或者，
参考信号的周期。
根据图3所示的流程可知，第一网元收集AI模型的候选训练数据，并根据第一信息对候选训练数据进行有效性的判定，可以理解为对所收集的候选训练数据进行筛选。在所收集的候选训练数据有效的情况下，第一网元将有效的候选训练数据发送给第二网元，用于第二网元进行AI模型的训练或更新，此时，该有效的候选训练数据即为训练数据。在第一网元收集的候选训练数据无效的情况下，第一网元向第二网元指示本次候选训练数据的收集无效。
可以理解的是，一次收集的候选训练数据无效，也代表此次未收集到训练数据。一次收集的候选训练数据有效，也代表此次收集到训练数据，此时，该有效的候选训练数据即成为训练数据，在本文中称为第一训练数据，由第一网元提供给第二网元。
应理解，在本申请各实施例中，“第一网元重新收集AI模型的训练数据”，也表示“第一网元重新收集AI模型的候选训练数据”。因为如果第一网元未收集到训练数据，才会尝试重新收集，而重新收集的目的在于期望收集到AI模型的训练数据，但是收集到AI模型的训练数据的过程是先收集候选训练数据，再从候选训练数据中筛选训练数据。
第二网元在接收到第一网元的第二信息之后，若根据第二信息确定第一网元的本次收集无效，在一种可能的情况下，第二网元确定需要重新收集AI模型的训练数据。
作为一个示例，第二网元向第一网元发送第三信息，第三信息指示第一网元重新收集AI模型的训练数据。可选地，在一个示例中，第三信息指示有效性判定的最大次数k。可选的，每次有效性判定都包括一批新的训练数据，即，一个新的训练数据集的收集，因而有效性判定的最大次数，也可称为执 行训练数据集收集的最大次数。
在此示例中，第二网元在接收到第一网元指示第一网元收集的候选训练数据无效的情况下，第二网元向第一网元发送第三信息，以指示第一网元重新收集AI模型的训练数据，同时，第三信息指示有效性判定的最大次数k，k为正整数。在第一次有效性判定所对应的候选训练数据无效的情况下，第一网元根据第三信息重新收集或继续收集候选训练数据。重新或继续收集可能涉及到多次，在每完成一次重新或继续收集之后，则进行一次有效性判定。如果判定结果为无效，则第一网元可以继续进行下一次重新收集，以及下一次的有效性判定，直至达到有效性判定的最大次数k。如果第1次有效性判定到第k-1次有效性判定的判定结果都为无效，且第k次有效性判定的判定结果依然为无效，第一网元将停止训练数据的收集。
在该示例中，第二网元在通过第三信息指示第一网元重新收集训练数据时，还可以向第一网元指示有效性判定的最大次数k。也即，该有效性判定的最大次数k的发送在第二网元确定需重新收集训练数据之后。该有效性判定的最大次数k可以包括在第三信息中，或者，通过除第三信息之外的其他信息携带。可选地，作为另一个示例，如步骤501中的描述，第二网元在向第一网元发送的第一信息中指示有效性判定的最大次数k。第二网元通过向第一网元指示有效性判定的最大次数k，对第一网元重新收集候选训练数据的过程进行了约束，使得第一网元在未收集到训练数据(也即有效的候选训练数据)的情况下，不会陷入无时间限制的重新收集的循环，而是在达到有效性判定的最大次数k之后，无论是否收集到训练数据，均停止收集。
在超过有效性判定的最大次数k之前，若第一网元根据第一信息确定第j次有效性判定的判定结果为有效，也即第j次有效性判定所针对的候选训练数据中包含有效的候选训练数据，则第一网元向第二网元发送第四信息，第四信息包括第二训练数据，且第四信息指示第j次有效性判定的判定结果为有效，第二训练数据具体可以包括第j次有效性判定所针对的候选训练数据中的有效的候选训练数据，j小于或等于k，j为正整数。
应注意，第j次有效性判定可以认为是针对一个候选训练数据的集合进行有效性判定的，而该集合中所包含的所有候选训练数据即为第j次有效性判定所针对的训练数据。其中，第j次有效性判定所针对的训练数据不限于为第j次收集到的候选训练数据，也可以包含第j次收集之前的一次或多次收集获得的候选训练数据，不作限定。
可选地，若第j次有效性判定的判定结果为有效，第一网元向第二网元发送第二训练数据以及信息a，信息a指示此次收集有效。
此外，在本申请各实施例中，有效性判定的最大次数k对应一个起始时刻，该起始时刻应理解为该有效性判定的最大次数k对应的训练数据的收集过程的起始时刻。示例性地，该起始时刻可以为第一网元接收第一信息或第三信息的时刻。相当于，第一网元从接收到第一信息或第三信息的时刻开始收集AI模型的训练数据。可选地，该收集过程的结束时刻是不确定的，例如，在超过有效性判定的最大次数k之前，若第j次有效性判定的判定结果为有效，则该收集过程结束，第一网元向第二网元发送有效的候选训练数据(也即第一训练数据)，j小于或等于k，j为正整数。但是，如果从第一次有效性判定直至第k次有效性判定，判定结果均为无效，则确定第k次有效性判定的判定结果为无效的时刻，为该收集过程的结束时刻。
本申请提供的获取训练数据的方法，训练数据的收集网元通过对收集到的候选训练数据进行有效性判定，并向AI模型的训练网元提供有效的候选训练数据，可以保证收集网元只向训练网元提供符合要求的训练数据，由于不符合要求的训练数据在收集网元侧被过滤掉，省去了无效的训练数据的交互，不仅节省了空口资源，也避免了对训练网元侧的训练数据集的污染，以及避免由此带来的其它不利影响。
以上对AI模型中获取训练数据的方法的主要流程进行了详细说明，下面针对AI模型应用于不同的场景中时，该获取训练数据的方法进行示例说明。
应用场景1
基于AI模型的信道状态信息(channel state information,CSI)反馈或CSI预测。
示例性地，应用场景1中的AI模型的训练或更新部署在接入网设备侧。接入网设备向UE发送下行参考信号，以便于UE通过测量下行参考信号获得测量结果，该测量结果即为候选训练数据。UE对 获得的候选训练数据进行有效性判定，并将有效的候选训练数据提供给接入网设备侧，用于AI模型的训练或更新。示例性地，在该应用场景下，下行参考信号具体可以为CSI-RS。UE向接入网设备提供的有效的候选训练数据为AI模型的标签，具体为CSI。
在很多应用场景下，接入网侧需要获取下行CSI，用于决定调度UE的下行数据信道的资源、调制编码方案(modulation and coding scheme,MCS)和预编码等配置中的一项或多项。在时分双工(time division duplex,TDD)系统中，由于上行信道和下行信道的互易性，接入网设备可以通过测量上行参考信号获得上行CSI，进而推测出下行CSI，例如，将上行CSI作为下行CSI。在频分双工(frequency division duplex,FDD)系统中，上行信道和上行信道的互易性无法保证，下行CSI是UE测量下行参考信号获得的，例如，UE测量CSI-RS或者同步信号和物理广播信道块(synchronizing signal and physical broadcast channel block,SSB)等信号获得下行CSI。UE按照协议预定义或者接入网设备预配置的方式生成CSI报告，并通过CSI报告将下行CSI反馈给接入网设备，使得接入网设备获得下行CSI。
参见图4，图4为一种基于AI模型的CSI反馈机制的示意图。如图4所示，自适应编码(auto encoder,AE)模型由编码器(encoder)和解码器(decoder)两个子模型构成，AE泛指由两个子模型构成的网络结构。AE模型也可以称为双边模型，或者双端模型或协作模型。AE的编码器和解码器通常是共同训练的，可以互相匹配使用。CSI反馈可以基于AE的AI模型实现。例如，UE侧测量基站发送的下行参考信号，获得测量的CSI。UE通过对编码器对测量获得的CSI进行压缩和量化，并向基站反馈经过压缩和量化后的信息，如图3中所示的“反馈的CSI的信息”。基站通过解码器对该“反馈的CSI的信息”进行恢复，获得恢复得到的CSI。对于基站而言，解码器的输入是UE反馈的CSI的信息，而解码器的训练获得需要UE测量获得的CSI作为恢复出的CSI的真值(也或者说标签)。
在应用场景1中，接入网设备侧部署的AI模型，可以如图4所示的解码器。
参见图5，图5为本申请提供的基于AI模型的CSI反馈中获取训练数据的示例。
501、可选地，接入网设备确定需要收集AI模型的训练数据。
502、接入网设备向UE发送第一信息，第一信息用于UE收集的候选训练数据的有效性的判定。可选地，有效性的判定结果可以为有效或无效。
示例性地，第一信息指示用于UE收集的候选训练数据的有效性的判定的约束条件。
关于第一信息以及约束条件等可以参考步骤310中的相关说明，这里不予赘述。
在应用场景1中，作为一个示例，测量结果的质量指标可以为训练数据的SINR和训练数据的数量。第一信息指示SINR的门限Q和训练数据的数量的门限N，而SINR的判定准则和训练数据的数量的判定准则(例如，SINR大于或等于Q，且训练数据的数量大于或等于N)可以是协议预定义的。作为另一个示例，第一信息指示SINR的门限Q和训练数据的数量的门限N，以及SINR的判定准则和训练数据的数量的判定准则。例如，第一信息指示门限Q和门限N，且第一信息包含用于指示判定准则的信息域。示例性地，该信息域包括1个比特，该1个比特对应SINR的判定准则和训练数据的数量的判定准则，例如，该1个比特的取值为“1”表示“训练数据的SINR大于或等于Q，并且，训练数据的数量大于或等于N”，该1个比特的取值为“0”表示“训练数据的SINR大于Q且训练数据的数量大于N”。示例性地，该信息域包括2个比特b1b0，其中b1对应SINR的判定准则，b0对应训练数据的数量的判定准则。例如，当b1的取值为1，表示“训练数据的SINR大于或等于Q”，当b1的取值为0，表示“训练数据的SINR小于Q”；b0指示训练数据的数量的判定准则也是类似的，不再赘述。作为再一个示例，第一信息指示训练数据的SINR的门限Q和训练数据的数量的门限N，此外，第一信息部分质量指标的判定准则，另一部分质量指标的判定准则是由协议预定义的。例如，第一信息指示门限Q和门限N，此外，第一信息还包括1个比特的信息域，当该1比特的取值为1时，表示“SINR大于或等于Q”，当该1比特的取值为0时，表示“SINR小于Q”；其中，训练数据的数量的判定准则由协议预定义，例如为“训练数据的数量至少为N个”。应理解，以上实现仅为第一信息用于确定约束条件的示例，不作限定。
示例性地，N可以为AI模型训练时batch的整数倍或者AI模型收敛所需的训练数据的数量。
503、接入网设备向UE发送参考信号。
UE通过对接入网设备的参考信号进行测量，获得一个或多个测量结果。在本申请中，测量结果也可以替换表达为参考信号的测量结果，或者信道测量结果。该替换表达也适用于其它应用场景下的实 施例中，下文不作重复说明。
可选地，测量结果包括信道响应，比如信道响应矩阵。
此外，可选地，UE可以通过一次测量，获得一个测量结果，此时，候选训练数据包括该一个测量结果；可选地，UE通过多次测量，获得多个测量结果，此时，候选训练数据包括该多个测量结果。
示例性地，测量结果的质量指标可以包括但不限于如下一项或多项：首径功率、首径到达时延、定时误差组(timing error group,TEG)、时域采样点的平均功率、天线端口之间的相位差、全带或子带的等效SINR、全带或子带的干扰水平、视距(line of light,LOS)概率、站间同步误差、或，测量结果置信度等中的一项或多项，不作限定。该指标质量适用于应用场景1或者下文介绍的其它应用场景中，不作限定。应理解，这些质量指标可以对参考信号的测量结果进行相应的处理来获得。具体处理的过程在此不予限定，比如，可以为已知的或未来的一些处理。
应理解，接入网设备向UE发送参考信号之前，接入网设备还会向UE发送参考信号对应的空口传输配置信息。空口传输配置信息指示接入网设备发送参考信号的相关空口配置，示例性地，该空口传输配置信息可以包括但不限于为参考信号的发送功率、接入网设备发送参考信号时所使用的天线端口数量、参考信号的频带宽度、参考信号的频域密度以及参考信号的周期等信息中的一项或多项。本领域技术人员可以理解，空口传输配置信息还可能包括其它的相关信息，这里不作一一罗列。
在该场景下，UE收集的候选训练数据为测量来自于接入网设备的参考信号获得的一个或多个测量结果，或者说一个或多个信道测量结果。
示例性地，以5G系统为例，该参考信号可以为信道状态信息-参考信号(channel state information,CSI-RS)。
504、UE根据第一信息对所收集的候选训练数据进行有效性的判定。
示例性地，假设候选训练数据为UE通过测量参考信号获得的多个测量结果，该多个测量结果即为候选训练数据。UE根据约束条件判定该多个测量结果中是否包含有效的候选训练数据(或者说有效的测量结果)。如上文的示例，若约束条件为“质量指标(例如SINR)大于或等于门限Q，且符合质量指标大于或等于门限Q的候选训练数据的数量至少为N个”，则UE判定收集到的多个测量结果中是否包含质量等于或大于门限Q的测量结果。为了描述上的简洁，以下将质量指标等于或大于门限Q的测量结果记作测量结果1。若UE确定收集的多个测量结果中包含测量结果1，还需要判断测量结果1的数量是否达到N个。若根据约束条件确定收集到有效的测量结果，则UE判定此次收集的候选训练数据有效，其中，有效的候选训练数据(即第一训练数据，下文有时也称为有效数据)即为满足约束条件的这部分测量结果。例如，如果测量结果1的数量为P个，P为大于或等于N的整数，则该P个测量结果1即为此次收集的有效数据，也即第一训练数据。
反之，如果UE判定收集到的多个测量结果中不包含满足约束条件的测量结果，例如，收集到的多个测量结果中包含SINR等于或大于门限Q的测量结果1，但是测量结果1的数量不足N个；或者，收集到的多个测量结果的SINR均小于门限Q，在此情况下，UE确定此次收集的候选训练数据无效。
505、UE根据所收集的候选训练数据的有效性的判定，向接入网设备发送第二信息，第二信息指示有效性的判定结果。
在一种可能的情况下，第二信息指示UE所收集的候选训练数据有效。在一个示例中，第二信息可以为所收集的有效候选训练数据本身，例如上文示例中的P个测量结果1。在该示例中，该P个测量结果1既是有效候选训练数据，同时该P个测量结果1也隐含指示UE所收集的候选训练数据有效。在另一个示例中，UE发送第二信息和有效的候选训练数据。在该示例中，第二信息指示UE所收集的候选训练数据有效，例如，第二信息可以包含1个比特，当该1比特的取值为“1”时，表示UE收集的候选训练数据有效。此外，UE向接入网设备发送有效的候选训练数据。相比之下，前一个示例在能够指示UE所收集的候选训练数据有效的前提下，能进一步节省信令开销。
在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。作为一个示例，第二信息可以包含1个比特，当该1比特的取值为“0”时，表示UE收集的候选训练数据无效。
示例性地，第二信息可以通过上行控制信息(uplink control information,UCI)信令携带，例如，UCI中包含1比特的信息，该1比特用于指示UE收集的候选训练数据为有效或无效。可选地，如上文的一个示例，若UE通过有效的候选训练数据隐含指示所收集的候选训练数据有效，该有效的候选 训练数据也可以在UCI中发送，不作限定。
506、接入网设备根据第二信息确定UE的候选训练数据的收集是否有效。
和步骤505中的判定结果相对应，在一种可能的情况下，第二信息指示UE所收集的候选训练数据有效，在此情况下，接入网设备还从UE获取UE收集的有效的候选训练数据。进一步地，接入网设备根据有效的候选训练数据进行AI模型的训练或更新，如步骤507。
507、接入网设备进行AI模型的训练，以获得该AI模型或对该AI模型进行更新。
在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。在此情况下的一种可能的实现中，接入网设备保持原有的AI模型的CSI反馈，或者切换到非AI模型的CSI反馈。作为一个示例，保持原有的AI模型可以是针对接入网设备上已经部署有训练好的AI模型，而此次训练数据的收集是基于更新AI模型的目的场景；切换到非AI模型可以是针对接入网设备上还没有训练好的AI模型，而此次训练数据的收集是为了训练获得AI模型的场景，在此场景下，若此次收集没有获得有效的候选训练数据，接入网设备可以切换到非AI模式的CSI反馈。这两种可能的情况如步骤508。
508、接入网设备基于原有的AI模型或切换到非AI模型进行CSI反馈。
在接入网设备执行步骤507或步骤508的情况下，一次训练数据的收集流程结束。
可选地，在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效，接入网设备获取到第二信息后，确定重新收集训练数据，如步骤509-510。
509、接入网设备确定重新收集AI模型的训练数据。
510、接入网设备向UE发送第三信息，第三信息指示UE重新收集AI模型的训练数据。
可选地，在一种可能的实现中，第三信息还指示有有效性判定的最大次数k，k为正整数。可选地，在另一种可能的实现中，有效性判定的最大次数k也可以由第一信息指示，不作限定。这两种实现在图3流程中已经详细说明，这里不再赘述。
可选地，在重新收集AI模型的训练数据的情况下，接入网设备可以更新空口传输配置。相应地，UE基于更新后的空口传输配置，重新收集AI模型的候选训练数据。
511、可选地，接入网设备向UE发送空口传输配置信息，该空口传输配置信息指示更新后的空口传输配置。
上文已经介绍过空口传输配置信息，如果空口传输配置有更新，步骤511中的空口传输配置信息指示更新后的空口传输配置。示例性地，空口传输配置的更新可以包括参考信号的发送功率、接入网设备发送参考信号时所使用的天线端口数量、参考信号的频带宽度、参考信号的频域密度、参考信号的周期等的更新等，不作限定。例如，空口传输配置的更新包括参考信号的发送功率增大、参考信号的频域密度增大，则接入网设备以更大的发送功率以及更大的频域密度向UE发送参考信号，以尝试让UE获得符合满足约束条件的候选训练数据。通过对空口传输配置的更新，可以保障空口基于AI模型的CSI反馈的精度性能。
当然，也可能在重新收集AI模型的训练数据时，不对原有的空口传输配置作更新。在此情况下，UE在原有的空口传输配置下重新收集候选训练数据，并基于约束条件对重新收集到的候选训练数据进行有效性的判定，并向接入网设备指示有效性的判定结果。
512、可选地，UE重新收集AI模型的训练数据。
应理解，在重新收集训练数据的流程中，重新收集的候选训练数据的有效性的判定，以及判定结果的指示与图5中的上述流程是类似的，不再赘述。需要理解的是，在重新收集训练数据的过程中，UE受到有效性判定的最大次数k的约束。
可选地，有效性判定的最大次数k可以是接入网设备根据训练数据收集的紧迫性确定的，示例性地，紧迫性可以是指AI模型的上一次更新至今的时间。例如，若AI模型的上一次更新至今的时间间隔较大，超过了某一个阈值，则认为AI模型的更新需求比较紧迫，因为时间间隔越大，意味着信道环境发生变化的可能性越大，代表AI模型对当前信道环境的匹配程度可能降低，因此更新需求则越紧迫。此时，有效性判定的最大次数k可以相应设置大一些，以便在一次无效收集之后，期望通过多次重新收集获得有效的候选训练数据。若AI模型的上一次更新至今的时间间隔很小，例如低于某个阈值，则认为紧迫性相对不足，有效性判定的最大次数k可以设置的小一些。可选地，紧迫性的判定准则也可以有其它实现，不作限定。
可见，将本申请提供的AI模型训练中用于获取训练数据的方法应用于基于AI模型的CSI反馈或CSI预测的场景下，可以减少AI模型训练流程中空口资源的浪费。此外，也避免了UE将无效的训练数据发送给接入网设备，对AI模型的训练数据集造成污染，从而影响AI模型的训练或更新，导致增益评估不准。
应用场景2
基于AI模型的定位场景。
由于上行定位和下行定位的不同，下面将对本申请的技术方案在上行定位和下行定位中的应用分别进行说明。
1、在上行定位中的应用。
参见图6，图6为本申请提供的技术方案在基于AI模型的上行定位场景的示意图。如图6，在上行定位中，AI模型的训练或更新部署在网络侧执行，以5G系统为例，AI模型可以部署在核心网的定位设备，例如LMF网元。其中，对于上行定位而言，AI模型的输入为一个或多个探测参考信号对应的一个或多个信道响应(或者说信道测量结果)，AI模型的输出为UE的位置。该一个或多个探测参考信号的发射端，如UE，可以为一个或多个，接收端，如接入网设备，也可以为一个或多个。
如图6所示，在上行定位中，定位设备对用于定位的AI模型进行训练时，从接入网侧获取一个接入网设备通过测量多个探测参考信号或多个接入网设备中的每个接入网设备分别通过测量一个或多个探测参考信号获得的多个测量结果，以及，第三网元的位置信息。其中，前述多个探测参考信号可以包括来自一个第三网元的多个探测参考信号，或，包括来自多个第三网元中每个第三网元的一个或多个探测参考信号。该一个第三网元的发送多个探测参考信号的不同时刻的位置信息，或，该多个第三网元在一个或多个时刻各自发送一个或多个探测参考信号时的位置信息，用于作为AI模型输出的位置信息的真值(即，标签)。
参见图7，图7为本申请提供的基于AI模型的上行定位中获取训练数据的示例。
701、可选地，定位设备确定需要收集AI模型的训练数据。
702、定位设备向接入网发送第一信息，第一信息用于接入网设备收集的候选训练数据的有效性的判定。可选地，判定结果可以为有效或无效。
示例性地，第一信息可以由定位设备和接入网设备之间的接口消息承载。以5G系统作为示例，若定位设备为LMF，接入网设备为gNB，则LMF和gNB之间的第一信息可以包含在NRPPa消息中。
703、接入网设备向第三网元发送空口传输配置信息(如空口传输配置信息#1)，空口传输配置信息指示第三网元发送探测参考信号时的空口传输配置。
在该实施例中，第三网元为可以提供自身位置信息的网元。在一个示例中，第三网元可以为位置参考设备。其中，位置参考设备可以视作一种特殊的网元，一般可以由网络厂商对其进行配置，例如网络厂商可以配置该位置参考设备的位置、发送能力、接收能力以及处理能力等中的一项或多项。位置参考设备可以向接入网设备提供其位置信息。示例性地，第三网元可以为参考UE，或者自动导航小车(automated guided vehicle,AGV)。在另一个示例中，第三网元也可以为普通UE。这里，普通UE是相对于位置参考设备而言的。普通UE可以通过一些定位方法获得自身的位置信息之后，将该位置信息提供该接入网设备。
704、接入网设备测量来自于第三网元的探测参考信号，获得一个或多个测量结果。
以5G系统为例，第三网元发送的探测参考信号可以为SRS(sounding reference signal)。
在步骤704中，接入网设备测量第三网元发送的探测参考信号，获得一个或多个测量结果，该一个或多个测量结果与第三网元的位置信息具有对应关系。在一个示例中，第三网元在位置1发送探测参考信号，接入网设备通过测量探测参考信号获得测量结果1，测量结果1对应位置1。第三网元在位置2发送探测参考信号，接入网设备通过测量探测参考信号获得测量结果2，测量结果2对应位置2。可选地，在另一个实例中，第三网元的绝对位置没有发生变化，但是第三网元的周围环境在不同的时间发生变化，接入网设备测量第三网元在不同的时间发送的探测参考信号，获得的测量结果也可能发生了变化。例如，接入网设备在时间1获得的测量结果1对应第三网元的位置1，在时间2获得的测量结果2对应第三网元的位置1。可选地，在再一个示例中，该实施例中的第三网元可以是多个。接入网设备分别测量来自于多个第三网元的探测参考信号，获得多个测量结果。也即，该多个测量结果中的 每个测量结果对应该多个第三网元中的一个第三网元的位置。相应地，在步骤705中，该多个第三网元分别向接入网设备或定位设备提供各自的位置信息。
705、第三网元提供自身的位置信息。
在该实施例中，以一个第三网元作为示例进行说明。一个位置信息对应一个或多个接入网设备对该第三网元在该位置信息所对应的位置所发送的一个或多个探测参考信号测量获得的一个或多个测量结果。在上行定位场景中，AI模型的候选训练数据为该一个或多个测量结果以及该一个或多个测量结果对应的第三网元的位置信息。在一个或多个接入网设备判定所收集的候选训练数据(即该一个或多个测量结果)有效的情况下，该一个或多个接入网设备将有效的候选训练数据分别提供给定位设备。
可选的，该有效的候选训练数据对应的第三网元的位置信息可以由第三网元通过该一个或多个接入网设备中的至少一个提供给定位设备，如步骤705a所示。应理解，705a为步骤705的一种实现方式。其中，该第三网元的位置信息可以对该至少一个接入网设备可见，或者，不可见。
可选地，在一种实现中，第三网元直接向定位设备提供子帧的位置信息(未图示)。定位设备获取到来自于一个或多个接入网设备的有效的候选训练数据，以及与该有效的候选训练数据对应的位置信息。应理解，有效的候选训练数据为多个，第三网元的位置信息也为多个。定位设备确定有效的候选训练数据和位置信息之间的对应关系。定位设备将位置信息作为AI模型的标签，对AI模型进行训练或更新，也即新建过程的训练或更新过程的训练。应理解，在该实施例中，定位设备获得的AI模型的有效的候选训练数据(也即，第一训练数据)包括：接入网设备测量第三网元发送的探测参考信号获得的测量结果中符合约束条件的一个或多个测量结果，以及和每个测量测量结果对应的第三网元的位置信息。其中，第三网元的位置信息为AI模型的输出真值，即标签。
706、接入网设备根据第一信息对所收集的候选训练数据进行有效性的判定。
需要说明的是，在步骤706中，接入网设备具体是对候选训练数据中的测量结果进行有效性的判定。
示例性地，在应用场景2中，测量结果的质量指标可以包括但不限于为首径功率、首径到达时延、定时误差组(timing error group,TEG)、时域采样点的平均功率、天线端口之间的相位差、全带或子带的等效SINR、全带或子带的干扰水平信息、视距(line of light,LOS)概率、站间同步误差指示信息、测量结果置信度指示信息中的一项或多项。此外，在定位的应用场景下，标签的质量指标可以为不同样本的位置之间的距离。
示例性地，约束条件中的质量指标可以为SINR和训练数据的数量。作为一个示例，所述训练数据的数量的门限为N，N可以为batch的整数倍或AI模型收敛所需训练数据的最少数量。可选地，在上行定位场景下，接入网设备收集的AI模型的候选训练数据还包括标签，该标签为位置信息，示例性地，约束条件中的质量指标还可以包括标签的质量指标，例如，标签的质量指标可以为不同样本的位置之间距离等，对此不作限定。
关于有效性判定可以参考步骤504的说明，这里不再赘述。
707、接入网设备根据候选训练数据的有效性的判定结果，向定位设备发送第二信息，其中，第二信息指示有效性的判定结果。
示例性地，第二信息可以包含在接入网设备和定位设备之间的接口消息中。或者说，接入网设备向定位设备发送接口消息，接口消息中包含第二信息。
708、定位设备根据第二信息确定接入网设备收集的候选训练数据是否有效。
在一种可能的情况下，第二信息指示接入网设备收集的候选训练数据有效，在此情况下，定位设备获取接入网设备收集的有效的候选训练数据(即，第一训练数据)。这里，第一训练数据具体包括符合约束条件的一个或多个测量结果以及该一个或多个测量结果各自对应的第三网元的位置信息。进一步地，定位设备根据有效的候选训练数据进行AI模型的训练或更新，如步骤709。
709、定位设备进行AI模型的训练，以获得该AI模型或对该AI模型进行更新。
在另一种可能的情况下，第二信息指示接入网设备所收集的候选训练数据无效。在此情况下的一种可能的实现中，定位设备保持原有的AI模型，或者切换到非AI模型，如步骤710。
710、定位设备基于原有的AI模型或切换到非AI模型进行上行定位。
可选地，在另一种可能的情况下，第二信息指示接入网设备所收集的候选训练数据无效。在另一 种可能的实现中，定位设备确定重新收集AI模型的训练数据，在此情况下，还包括步骤711以及712。
711、定位设备确定重新收集训练数据。
712、定位设备向接入网设备发送第三信息，第三信息指示接入网设备重新收集AI模型的训练数据。
可选地，第三信息还指示有效性判定的最大次数k，k为正整数。可选地，有效性判定的最大次数k也可以由第一信息指示，可以参考图3所示流程中的相关说明，不再赘述。
在该应用场景下，作为一个示例，有效性判定的最大次数k可以由定位设备根据AI模型的训练数据需求的紧迫性来配置，这与应用场景1中是类似的。作为一个示例，紧迫性的判断准则可以是根据当前AI模型对第三网元的位置的估计结果误差或者AI模型的上次更新时间至今的时间间隔确定的。例如，若定位设备根据当前的AI模型对第三网元的位置的估计结果误差较大，例如，大于或等于某个设定的阈值，则可以判定为紧迫。此情况下，有效性判定的最大次数k可以设置的大一些；反之，如果基于当前的AI模型对第三网元的位置的估计结果误差较小，例如小于该设定的阈值，则可以判定为不紧迫。此情况下，有效性判定的最大次数k可以设置的小一些。其中，AI模型的估计结果误差的判断是通过将上一次AI模型训练所收集到的训练数据拆分为训练集和验证集实现的。由于训练集的误差已经很低，因此将验证集的估计结果误差作为AI模型是否严重失效的判断准则。此外，也可以根据AI模型的上次更新时间至今的时间间隔设定，可以参考应用场景1中的解释说明，不予赘述。
可选地，定位设备在确定重新收集AI模型的训练数据的情况下，可以指示接入网设备更新接入网设备收集训练数据时的空口传输配置，如步骤713。
713、接入网设备向第三网元发送空口传输配置信息(如空口传输配置信息#2)，空口传输配置信息指示更新的空口传输配置。
应理解，步骤713中空口传输配置的更新，是相对于步骤703中的空口传输配置而言的更新。示例性地，该更新包括但不限于：增大探测参考信号的发送功率、增大探测参考信号的频域密度等。应理解，更新空口传输配置的目的在于，接入网设备尝试收集到符合约束条件的候选训练数据，并提供给定位设备。
714、接入网设备重新收集AI模型的训练数据。
可见，将本申请提供的AI模型训练中获取训练数据的方法应用于基于AI模型的上行定位的场景下，可以减少AI模型训练流程中空口资源的浪费。此外，也避免了接入网设备将无效的训练数据发送给定位设备，对AI模型的训练数据集造成污染，从而影响AI模型的训练或更新，导致增益评估不准。
应理解，在上行定位场景中，接入网设备为训练数据的收集网元的一个示例，定位设备为AI模型的训练网元的一个示例。
2、在下行定位中的应用。
参见图8，图8为本申请提供的技术方案在基于AI模型的下行定位场景的示意图。如图8，在下行定位中，AI模型推理部署在UE侧，但AI模型的训练部署在网络侧的定位设备，例如LMF网元。定位设备上部署的AI模型以UE测量参考信号得到的对应的信道响应为输入，UE的位置为输出。以5G系统为例，该参考信号可以为定位参考信号，可以由一个或多个基站(base station,BS)发送给UE。
参见图9，图9为本申请提供的基于AI模型的上行定位中获取训练数据的示例。
801、可选地，定位设备确定需要收集AI模型的训练数据。
该AI模型的训练数据可以来自一个UE对多个参考信号的测量，或，多个UE中每个UE分别对一个或多个参考信号的测量。其中，该多个参考信号可以来自一个或多个接入网设备。本实施例从定位设备和该一个UE或多个UE中的某一个UE之间的通信的视角进行描述。
802、定位设备向UE发送第一信息，第一信息用于UE收集的候选训练数据的有效性的判定。可选地，有效性的判定结果可以为有效或无效。
可选地，作为一个示例，定位设备通过接入网设备向UE发送第一信息，如图9中所示的步骤802a和步骤802b。作为另一个示例，定位设备也可以通过和UE之间的接口，直接向UE发送第一信息。定位设备向接入网设备发送信息#1，信息#1指示接入网设备向UE发送定位参考信号。可选地，信息#1也可以为第一信息。接入网设备基于信息#1的触发，向UE发送定位参考信号。图9中所示实现仅作为示例。
803、接入网设备向UE发送定位参考信号。
UE测量来自于接入网设备，或者，UE测量该接入网设备以及其他接入网设备，的定位参考信号，获得候选训练数据，具体为定位参考信号的一个或多个测量结果，以及该一个或多个测量结果对应的UE的位置信息。
应理解，接入网设备向UE发送PRS之前，还向UE发送PRS的空口传输配置信息，以指示PRS的空口传输配置。
以5G系统为例，接入网设备向UE发送的定位参考信号可以为PRS。候选训练数据为UE测量PRS获得的一个或多个测量结果以及该一个或多个测量结果对应的UE的位置信息。
804、UE根据第一信息判定所收集的候选训练数据的有效性。
示例性地，UE根据第一信息指示的约束条件，判定所收集的候选训练数据的有效性，这里具体是判定所述该一个或多个测量结果的有效性。与步骤504类似，可以参考步骤504理解，这里省略详细说明。另外，关于下行定位场景下，约束条件包含的质量指标的示例可以参考上行定位场景下的说明，这里不再赘述。
805、UE向定位设备发送第二信息，第二信息指示有效性的判定结果。
在一种可能的情况下，第二信息指示UE所收集的候选训练数据有效。在一个示例中，第二信息可以为UE所收集的候选训练数据中的第一训练数据。其中，第一训练数据包括UE的位置信息。或者说，第一训练数据具体为符合约束条件的测量结果及其对应的UE的位置信息。在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。
可选地，步骤805中，UE可以通过UE和定位设备之间的接口，直接向定位设备发送第二信息，如图9中所示。或者，UE也可以向接入网设备发送第二信息，再由接入网设备将第二信息发送给定位设备。或者，在UE所收集的候选训练数据有效的情况下，UE向接入网设备发送第二信息包含的部分信息，例如第一训练数据所包含的符合约束条件的测量结果(也即有效的测量结果)，并向定位设备发送自身的位置信息。接入网设备再将符合约束条件的测量结果发送给定位设备。由此定位设备获取到第一训练数据，其中，第一训练数据包括有效的测量结果及其对应的UE的位置信息，不作限定。
806、定位设备根据第二信息确定UE收集的候选训练数据的收集是否有效。
在一种可能的情况下，第二信息指示UE所收集的候选训练数据有效，在此情况下，第二信息可以包含第一训练数据，其中，第一训练数据包括UE的位置信息。进一步地，定位设备根据第一训练数据进行AI模型的训练或更新，如步骤807。
807、定位设备进行AI模型的训练，以获得该AI模型或对该AI模型进行更新。
在另一种可能的情况下，若第二信息指示UE所收集的候选训练数据无效。在此情况下的一种可能的实现中，接入网设备保持原有的AI模型的波束管理或者切换到非AI模型进行下行定位，如步骤808。
808、定位设备保持原有AI模型或切换到非AI模型进行定位。
可选地，在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。在一种可能的实现中，定位设备确定重新收集训练数据，如步骤809。
809、定位设备确定重新收集训练数据。
810、定位设备向UE发送第三信息，第三信息指示UE重新收集AI模型的训练数据。
可选地，第三信息指示还可以指示有效性判定的最大次数k。或者第一信息指示有效性判定的最大次数k。
可选地，在重新收集AI模型的训练数据的情况下，定位设备可以指示接入网设备更新空口传输配置。例如，定位设备向接入网设备发送信息#2，信息#2用于指示接入网设备重新收集训练数据。相应地，接入网设备向UE发送更新后的空口传输配置对应的空口传输配置信息，如步骤811。
811、接入网设备向UE发送空口传输配置信息，该空口传输配置信息指示更新后的空口传输配置。
UE基于更新后的空口传输配置，测量接入网设备发送的定位参考信号，以重新收集AI模型的训练数据。
812、UE重新收集AI模型的训练数据。
可选地，该实施例中的UE可以为位置参考设备，也可以普通UE，不作限定。位置参考设备或普 通UE可以参考步骤703中的说明，不再赘述。
此外，应理解，该实施例中的UE为训练数据的收集网元的一个示例，定位设备为AI模型的训练网元的一个示例。
可见，将本申请提供的AI模型训练中获取训练数据的方法应用于基于AI模型的下行定位的场景下，可以减少AI模型训练流程中空口资源的浪费。此外，也避免了终端设备(如位置参考设备)将无效的训练数据发送给定位设备，对AI模型的训练数据集造成污染，从而影响AI模型的训练或更新，导致增益评估不准。
应用场景3
基于AI模型的波束管理。
示例性地，应用场景3中的AI模型的训练部署在接入网设备侧。接入网设备需要获取UE侧通过测量参考信号获得的训练数据(例如，参考信号的一个或多个测量结果)，并将从UE侧获取的训练数据用于AI模型的训练或更新。在该应用场景下，AI模型的标签为最优的K个测量结果对应的参考信号的信息。可替换地，应用场景3中，最优的K个测量结果对应的参考信号的信息，也可以替换为该最优的K个测量结果对应K个波束的信息，示例性地，为K个波束ID。
可知，5G系统引入6GHz以上的高频段用于数据通信，相对于6GHz以下的中低频段，高频段频谱的连续可用带宽较大，中心频率较高，因而可以获得更大的传输速率和系统容量。但是由于高频信号(例如毫米波)穿透能力弱、路径衰落效应强，导致高频信号的传播距离受限，覆盖能力堪忧。得益于大规模天线技术，高频通信系统通常采用数量众多的天线做波束赋型，从而可以获取可观的波束增益来补偿高频传播特性导致的传播距离受限。但是，设计精准的波束赋型时，基站需要从终端获取准确的信道信息，获取如此大规模的天线阵列上的信道信息需要消耗庞大的空口开销，在实际系统中不可接受。人们在实验中发现，高频无线信道具有明显的稀疏性，即信道的主要能量集中在有限个数的径上，例如，信号的发射端和接收端之间存在无遮挡直射视距(line of sight,LoS)路径时，接收端和发射端之间的主要能量集中在直射视距路径上。当发射端和接收端之间存在非视距(non-line of sight,NLOS)的遮挡时，其主要能量多集中在反射一次即可到达的路径上。通常每条径有着不同的入射和出射角度，故高频通信系统的发射端和接收端只需将自己的波束方向对准信道主要径的入射角度和出射角度，即可获取大部分的信道传输能量完成通信。
对于一个高频通信系统，假设发射端总共有S根天线，接收端有R根天线，其形态可以包括线天线或者面阵天线。收发两端在自己的天线上乘以不同的预编码权值，对发送信号进行预编码，就可使得发出的信号具有波束赋型的效果。例如，对于下行信号传输模型：
Y＝VHWX+N
接收端的接收预编码矩阵为V，信道响应为H。发射端的发射预编码矩阵为W，发射信号为X，噪声为N。接收端接收到的信号为Y。发射端发射预编码矩阵的形式为W＝[W1,W2,…WM]，其中，Wi为发射端第i根天线上的预编码权值。类似地，接收端的接收预编码矩阵的形式为V＝[V1,V2,…VR]，其中，Vi为接收端第i根天线上的预编码权值。发射信号X经过W的预编码后得到的信号为WX,WX是发射端最终的发射信号。WX在空间中具有波束赋形的效果。根据X上所承载信息的差异性，可将WX分为参考信号和数据信号。波束管理过程中发送的一般为参考信号，可能的参考信号的种类包括SSB、CSI-RS、SRS、相位跟踪参考信号(phase-tracking reference signal,PTRS)、解调参考信号(demodulation reference signal,DMRS)等。
由于信道主径的角度可以分布在一个很宽的范围内，如0～360度，而每个预编码矩阵W在空间中只能覆盖一定的角度范围，对应一个赋型波束，故需要设计多个预编码矩阵W才能保证较好的信号覆盖效果。多个指向角度不同的预编码矩阵W构成一个码本。发射端和接收端都会维护一套自己的码本。在波束管理过程中，发射端和接收端通过遍历扫描自己的码本实现收发双端的角度对准。例如，发射端的码本中有64个预编码矩阵，分别对应64个赋型波束。接收端码本有4个预编码矩阵，分别对应4个赋型波束，则总共需要扫描256(64*4)次才能确定一对最优的收、发端赋型波束，扫描开销和时延非常大。
在一种基于AI模型的波束管理技术中，可以实现稀疏的波束扫描，大幅降低波束扫描的开销和时 延。对于一个高频通信系统，假设发射端总共有S根天线，接收端有R根天线，发射端的码本中有S个预编码矩阵(S个赋型波束)，接收端的码本中有R个预编码矩阵(R个赋型波束)。对于每个接收赋型波束来说，任何一个发射赋型波束都可能是与之构成一个收发波束对，故确定最优收发波束对的过程可以拆解成对于某个接收波束，进行发射端波束扫描确定匹配的最优发射波束，然后将此过程对剩下R-1个接收波束分别重复一遍即可确定全局最优的收发波束对。同样的，对于每个发射波束来说，任何一个接收波束都可能是与之构成一个收发波束对，故确定最优收发波束对的过程可以拆解成对于某个发射波束，进行接收端波束扫描确定匹配的最优发射波束，然后将此过程对剩下S-1个发射波束分别重复一遍即可确定全局最优的收发波束对。故下文以发射端波束扫描为例进行阐述。
参见图10，图10为基于AI辅助的稀疏波束扫描过程示意图。
假设发射端码本中的预编码矩阵对应第一值，如64，个赋型波束，传统方案需扫描全部的64个波束以确定最优波束，但AI辅助(也即，基于AI模型的)的稀疏波束扫描方案只需扫描码本中的部分波束，例如图10中填充标记的第二值，如16，个波束。从码本选择部分波束用于扫描可以有很多种选法，每一种被选出的波束组合称为一种稀疏波束图样。发射端用稀疏波束图样中的预编码矩阵进行预编码，发送参考信号。接收端将参考信号的测量结果输入用于AI波束预测的神经网络，神经网络输出K个波束的索引，也称为Top-K波束的索引。应注意，该K个波束是码本中所有64个赋型波束中的K个，而并非局限于稀疏波束图样包含的赋型波束中的K个。接收端将Top-K波束的索引反馈给发射端。可选的，当K＞1时，发射端仅扫描此K个赋型波束，发射经过波束赋型的参考信号，接收端采用能量检测的方法，测量此K个参考信号的能量并选择能量最强的作为最优波束。
下面介绍本申请提供的AI模型训练中获取训练数据的方法在基于稀疏波束扫描技术中的波束管理中的应用。在该应用中，发射端为接入网设备，例如基站。接收端为UE。UE测量来自于接入网设备的参考信号，获得多个测量结果，并确定出TOP-K波束索引，并将该多个测量结果和TOP-K波束索引反馈给接入网设备。
参见图11，图11为本申请提供的基于AI模型的波束管理中获取训练数据的示例。
901、可选地，接入网设备确定需要收集AI模型的训练数据。
902、接入网设备向UE发送第一信息，第一信息用于UE收集的候选训练数据的有效性的判定。可选地，有效性的判定结果可以为有效或无效。
903、接入网设备向UE发送多个参考信号。
UE测量来自于接入网设备的多个参考信号，获得多个测量结果，即候选训练数据。可选的，该多个参考信号对应前述第一值，如64，个赋型波束。
可选地，在应用场景3中，参考信号可以为CSI-RS和/或SSB。在该示例中，CSI-RS和/或SSB用于UE进行信道测量。
904、UE根据第一信息判定所收集的候选训练数据的有效性。
905、UE向接入网设备发送第二信息，第二信息指示有效性的判定结果。
在一种可能的情况下，第二信息指示UE所收集的候选训练数据有效。在一个示例中，第二信息可以为第一训练数据，其中，第一训练数据包括所述一个或多个测量结果中的K个最优的测量结果对应的参考信号的信息，K为大于或等于1的整数。示例性地，该K个最优的测量结果对应的参考信号的信息，可以理解为图10中的TOP-K波束的索引。另一个示例中，第二信息可以为第一训练数据，其中，第一训练数据包括所述多个测量结果中的L个测量结果，以及该L个测量结果对应的L个参考信号的信息。其中，该L个测量结果为有效的测量结果，也即符合约束条件的测量结果。在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。
在该应用场景中，参考信号也可以替换为“波束”，不作限定。
906、接入网设备根据第二信息确定UE收集的候选训练数据是否有效。
在一种可能的情况下，第二信息指示UE所收集的候选训练数据有效，在此情况下，接入网设备从UE获取第一训练数据。进一步地，接入网设备根据第一训练数据进行AI模型的训练或更新，如步骤907。
907、接入网设备根据第一训练数据进行AI模型的训练，以获得该AI模型或对该AI模型进行更新。
在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。在此情况下的一种可能的实现中，接入网设备保持原有的AI模型的波束管理或者切换到非AI模型进行波束管理，如步骤908。
908、接入网设备保持原有的AI模型不变，或者切换到非AI模型进行波束管理。
可选地，在另一种可能的情况下，第二信息指示UE所收集的候选训练数据无效。在一种可能的实现中，接入网设备确定重新收集训练数据，如步骤909。
909、接入网设备确定重新收集训练数据。
910、接入网设备向UE发送第三信息，第三信息指示UE重新收集AI模型的训练数据。
可选地，第三信息还指示有效性判定的最大次数k。或者，有效性判定的最大次数k也可以由第一信息指示，不作限定。
可选地，在重新收集AI模型的训练数据的情况下，接入网设备可以更新空口传输配置。相应地，UE基于更新后的空口传输配置，重新收集AI模型的候选训练数据。
911、接入网设备向UE发送空口传输配置信息，该空口传输配置信息指示更新后的空口传输配置。
912、UE重新收集AI模型的训练数据。
与上述两个应用场景类似，在重新收集候选训练数据的过程中，UE受到有效性判定的最大次数k的限定。可选地，有效性判定的最大次数k可以是接入网设备根据训练数据收集的紧迫性确定的，示例性地，紧迫性的判断准则可以是指AI模型的上一次更新至今的时间，或者是接入网设备根据当前的AI模型对最优测量结果(或者说，最优波束)的估计结果误差设定的。其中，AI模型的估计结果误差的判断是通过将上一次AI模型训练所收集到的训练数据拆分为训练集和验证集实现的。由于训练集的误差已经很低，因此将验证集的估计结果误差作为AI模型是否失效严重的判断准则。例如，若接入网设备根据当前的AI模型对UE接收参考信号时的最优测量结果对应的参考信号的信息的预测结果误差较大，例如，大于或等于某个设定的阈值，则可以判定为紧迫。此情况下，有效性判定的最大次数k可以设置的大一些；反之，如果基于当前的AI模型对UE接收参考信号时的最优测量结果对应的参考信号的信息的预测结果误差较小，例如小于该设定的阈值，则可以判定为不紧迫。此情况下，有效性判定的最大次数k可以设置的小一些。此外，也可以根据AI模型的上次更新时间至今的时间间隔设定，可以参考应用场景1中的解释说明，不予赘述。
可见，将本申请提供的AI模型训练中用于获取训练数据的方法应用于基于AI模型的波束管理的场景下，可以减少AI模型训练流程中空口资源的浪费。此外，也避免了UE将无效的训练数据发送给接入网设备，而对AI模型的训练数据集造成污染。
以上图3至图11中涉及到方法流程的示意图中，各步骤的编号仅仅是为了清楚地描述本申请的技术方案，不应该对方法的具体实现构成限定。这些步骤可以扩展为更多的步骤，或者也可以合并为更少的步骤，取决于不同的具体实现，不作限定。此外，图3至图11中的虚线所示的步骤表示为可选步骤。
以上对本申请提供的AI模型训练中获取训练数据的方法进行了详细说明。基于相同的技术构思，参见图12，本申请提供了一种通信装置1000。
如图12，通信装置1000包括处理模块1001和通信模块1002。该通信装置1000可以是终端设备，也可以是应用于终端设备或者和终端设备匹配使用，能够实现终端设备侧执行的通信方法的通信装置，例如，芯片或电路；或者，该通信装置1000可以是网络设备，也可以是应用于网络设备侧或者和网络设备侧匹配使用，能够实现网络设备侧执行的通信方法的通信装置，例如芯片或电路。示例性地，该网络设备侧例如可以为本申请方法实施例中的接入网设备或定位设备。
其中，通信模块也可以称为收发模块、收发器、收发机、或收发装置等。处理模块也可以称为处理器，处理单板，处理单元、或处理装置等。可选的，通信模块用于执行上述方法中终端设备侧或网络设备侧的发送操作和接收操作，可以将通信模块中用于实现接收功能的器件视为接收单元，将通信模块中用于实现发送功能的器件视为发送单元，即通信模块包括接收单元和发送单元。
该通信装置1000应用于终端设备时，处理模块1001可用于实现图3～图11所述各实施例中所述终端设备的处理功能，通信模块1002可用于实现图3～图11所述各实施例中所述终端设备的收发功能。或者也可以参照发明内容中第三方面以及第三方面中可能的设计理解该通信装置。
该通信装置1000应用于网络设备时，处理模块1001可用于实现图3～图11所述各实施例中网络 设备(例如，接入网设备或定位设备)的处理功能，通信模块1002可用于实现图3～图11所述各实施例中网络设备的收发功能。或者也可以参照发明内容中第四方面以及第四方面中可能的设计理解该通信装置。
需要说明的是，图3中所示的第一网元或第二网元具体为终端设备或网络设备(例如接入网设备或定位设备)已经在前述方法实施例中，针对各种不同的应用场景作了详细说明，可以参考具体的实施例来理解第一网元为终端设备或网络设备，这里不再赘述。
此外需要说明的是，前述通信模块和/或处理模块可通过虚拟模块实现，例如处理模块可通过软件功能单元或虚拟装置实现，通信模块可以通过软件功能或虚拟装置实现。或者，处理模块或通信模块也可以通过实体装置实现，例如若该装置采用芯片/芯片电路实现，所述通信模块可以是输入输出电路和/或通信接口，执行输入操作(对应前述接收操作)、输出操作(对应前述发送操作)；处理模块为集成的处理器或者微处理器或者集成电路。
本申请中对模块的划分是示意性的，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，另外，在本申请各个示例中的各功能模块可以集成在一个处理器中，也可以是单独物理存在，也可以两个或两个以上模块集成在一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。
基于相同的技术构思，参见图13，本申请还提供了一种通信装置1100。可选地，该通信装置1100可以是芯片或者芯片系统。可选的，在本申请中芯片系统可以由芯片构成，也可以包含芯片和其他分立器件。
通信装置1100可用于实现前述示例描述的通信系统中任一网元的功能。通信装置1100可以包括至少一个处理器1110。可选的，该处理器1110与存储器耦合，存储器可以位于该装置之内，或，存储器可以和处理器集成在一起，或，存储器也可以位于该装置之外。例如，通信装置1100还可以包括至少一个存储器1120。存储器1120保存实施上述任一示例中必要计算机程序、计算机程序或指令和/或数据；处理器1110可能执行存储器1120中存储的计算机程序，完成上述任一示例中的方法。
通信装置1100中还可以包括通信接口1130，通信装置1100可以通过通信接口1130和其它设备进行信息交互。示例性的，所述通信接口1130可以是收发器、电路、总线、模块、管脚或其它类型的通信接口。当该通信装置1100为芯片类的装置或者电路时，该装置1100中的通信接口1130也可以是输入输出电路，可以输入信息(或称，接收信息)和输出信息(或称，发送信息)，处理器为集成的处理器或者微处理器或者集成电路或则逻辑电路，处理器可以根据输入信息确定输出信息。
本申请中的耦合是装置、单元或模块之间的间接耦合或通信连接，可以是电性，机械或其它的形式，用于装置、单元或模块之间的信息交互。处理器1110可能和存储器1120、通信接口1130协同操作。本申请中不限定上述处理器1110、存储器1120以及通信接口1130之间的具体连接介质。
可选的，如图13中所示，所述处理器1110、所述存储器1120以及所述通信接口1130之间通过总线1140相互连接。所述总线1140可以是外设部件互连标准(peripheral component interconnect,PCI)总线或扩展工业标准结构(extended industry standard architecture,EISA)总线等。所述总线可以分为地址总线、数据总线、控制总线等。为便于表示，图13中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
在本申请中，处理器可以是通用处理器、数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本申请中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者任何常规的处理器等。结合本申请所公开的方法的步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。
在本申请中，存储器可以是非易失性存储器，比如硬盘(hard disk drive,HDD)或固态硬盘(solid-state drive,SSD)等，还可以是易失性存储器(volatile memory)，例如随机存取存储器(random-access memory,RAM)。存储器是能够用于携带或存储具有指令或数据结构形式的期望的程序代码并能够由计算机存取的任何其他介质，但不限于此。本申请中的存储器还可以是电路或者其它任意能够实现存储功能的装置，用于存储程序指令和/或数据。
在一种可能的实施方式中，该通信装置1100可以应用于网络设备侧，例如本申请实施例中的接入 网设备或定位设备。具体地，通信装置1100可以是网络设备，也可以是能够支持网络设备实现上述涉及的任一示例中网络设备侧相应的功能的装置。存储器1120保存实现上述任一示例中的网络设备侧的功能的计算机程序(或指令)和/或数据。处理器1110可执行存储器1120存储的计算机程序，完成上述任一示例中网络设备侧执行的方法。应用于接入网设备时，该通信装置1100中的通信接口可用于与终端设备进行交互，向终端设备发送信息或者接收来自终端设备的信息；此外，可选地，该通信装置1000中的通信接口还可用于与核心网设备进行交互，例如与定位设备(例如LMF网元)进行交互，向定位设备发送信息或接收来自于定位设备的信息。
在另一种可能的实施方式中，该通信装置1100可以应用于终端设备，具体地，通信装置1100可以是终端设备，也可以是能够支持终端设备，实现上述涉及的任一示例中终端设备的功能的装置。存储器1120保存实现上述任一示例中的终端设备的功能的计算机程序(或指令)和/或数据。处理器1110可执行存储器1120存储的计算机程序，完成上述任一示例中终端设备执行的方法。应用于终端设备，该通信装置1100中的通信接口可用于与网络设备侧(例如，接入网设备)进行交互，向网络设备侧发送信息或者接收来自接入网设备的信息。
由于本示例提供的通信装置1100可应用于网络设备侧(例如接入网设备或定位设备)，完成上述网络设备侧执行的方法，或者应用于终端设备，完成终端设备执行的方法。因此其所能获得的技术效果可参考上述方法实施例中的说明，在此不再赘述。
基于以上示例，本申请提供了一种通信系统，包括网络设备和终端设备。在一个示例中，该通信系统包括接入网设备和终端设备。在另一个示例中，该通信系统包括定位设备、接入网设备和终端设备。其中，所述接入网设备和终端设备，或者所述定位设备、所述接入网设备和终端设备，可以实现图3～图11所示的示例中所提供的通信方法。
本申请提供的技术方案可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、终端设备、接入网设备或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(digital subscriber line，DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机可以存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质(例如，软盘、硬盘、磁带)、光介质(例如，数字视频光盘(digital video disc，DVD))、或者半导体介质等。
在本申请中，在无逻辑矛盾的前提下，各示例之间可以相互引用，例如方法实施例之间的方法和/或术语可以相互引用，例如装置实施例之间的功能和/或术语可以相互引用，例如装置示例和方法示例之间的功能和/或术语可以相互引用。
在本说明书中使用的术语“部件”、“模块”、“系统”等用于表示计算机相关的实体、硬件、固件、硬件和软件的组合、软件、或执行中的软件。例如，部件可以是但不限于，在处理器上运行的进程、处理器、对象、可执行文件、执行线程、程序和/或计算机。通过图示，在计算设备上运行的应用和计算设备都可以是部件。一个或多个部件可驻留在进程和/或执行线程中，部件可位于一个计算机上和/或分布在2个或更多个计算机之间。此外，这些部件可从在上面存储有各种数据结构的各种计算机可读介质执行。部件可例如根据具有一个或多个数据分组(例如来自与本地系统、分布式系统和/或网络间的另一部件交互的二个部件的数据，例如通过信号与其它系统交互的互联网)的信号通过本地和/或远程进程来通信。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统、装置和单元的 具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (42)

  1. 一种人工智能AI模型训练中用于获取训练数据的方法，其特征在于，所述方法由第一网元或用于第一网元的芯片执行，所述方法包括：
    接收来自于第二网元的第一信息，所述第一信息用于收集的候选训练数据的有效性的判定，所述有效性的判定结果包括有效或无效；
    收集所述AI模型的候选训练数据；
    根据所述候选训练数据和所述第一信息，向第二网元发送第二信息，所述第二信息指示所述有效性的判定结果。
  2. 根据权利要求1所述的方法，其特征在于，所述第二信息包括第一训练数据且所述第二信息指示所述收集的所述候选训练数据有效，所述第一训练数据为所述候选训练数据中的有效数据。
  3. 根据权利要求1所述的方法，其特征在于，所述第二信息指示所述收集的所述候选训练数据无效。
  4. 根据权利要求1至3中任一项所述的方法，其特征在于，所述第一信息用于所述收集的所述候选训练数据的有效性的判定的约束条件的确定。
  5. 根据权利要求4所述的方法，其特征在于，所述方法还包括：
    若确定所述候选训练数据中包含满足所述约束条件的第一训练数据，确定所述候选训练数据有效；或者，
    若确定所述候选训练数据中不包含满足所述约束条件的第一训练数据，确定所述候选训练数据无效。
  6. 根据权利要求3至5中任一项所述的方法，其特征在于，在所述收集的所述候选训练数据无效的情况下，所述方法还包括：
    接收来自于所述第二网元的第三信息，所述第三信息指示重新收集所述AI模型的候选训练数据。
  7. 根据权利要求6所述的方法，其特征在于，所述方法还包括：
    确定空口传输配置信息，所述空口传输配置信息对应更新的空口传输配置，所述空口传输配置信息指示基于所述更新的空口传输配置收集所述AI模型的候选训练数据；
    其中，所述更新的空口传输配置信息包括如下一项或多项的更新：
    参考信号的发送功率；
    参考信号使用的天线端口数；
    参考信号的频带宽度；
    参考信号的频域密度；或，
    参考信号的周期。
  8. 根据权利要求6或7所述的方法，其特征在于，所述第三信息还指示所述有效性的判定的最大次数k，k为正整数。
  9. 根据权利要求6或7所述的方法，其特征在于，所述第一信息还指示所述有效性的判定的最大次数k，k为正整数。
  10. 根据权利要求8或9所述的方法，其特征在于，所述方法还包括：
    基于所述更新的空口传输配置，收集所述AI模型的候选训练数据；
    若达到所述有效性的判定的最大次数k，且根据所述第一信息确定第k次有效性的判定结果为无效，停止收集所述AI模型的候选训练数据。
  11. 根据权利要求10所述的方法，其特征在于，所述方法还包括：
    在超过所述有效性的最大判定次数k之前，若根据所述第一信息确定第j次有效性的判定结果为有效，向所述第二网元发送第四信息，所述第四信息包括第二训练数据，且所述第四信息指示所述第j次有效性的判定结果为有效，所述第二训练数据包括所述第j次有效性的判定所针对的候选训练数据中的有效数据，j小于或等于k，j为正整数。
  12. 根据权利要求1至11中任一项所述的方法，其特征在于，所述收集所述AI模型的候选训练 数据，包括：
    测量来自于所述第二网元的参考信号，获得一个或多个测量结果，所述AI模型的候选训练数据包括所述一个或多个测量结果；或者，
    测量来自于第三网元的参考信号，获得一个或多个测量结果，所述AI模型的候选训练数据包括所述一个或多个测量结果。
  13. 根据权利要求12所述的方法，其特征在于，所述第一网元为终端设备或用于所述终端设备的芯片，所述第二网元为接入网设备或用于所述接入网设备的芯片；
    所述第一网元测量来自于所述第二网元的参考信号，获得所述一个或多个测量结果。
  14. 根据权利要求13所述的方法，其特征在于，所述第一训练数据还包括所述一个或多个测量结果中的K个最优的测量结果对应的参考信号的信息或波束信息，K为大于或等于1的整数。
  15. 根据权利要求12所述的方法，其特征在于，所述第一网元为接入网设备或用于所述接入网设备的芯片，所述第二网元为定位设备或用于所述定位设备的芯片；
    所述第一网元测量来自于所述第三网元的探测参考信号，获得所述一个或多个测量结果；
    以及，所述第一训练数据还包括所述第三网元的一个或多个位置信息。
  16. 根据权利要求12所述的方法，其特征在于，所述第一网元为终端设备或用于所述终端设备的芯片，所述第二网元为定位设备或用于所述定位设备的芯片；
    所述第一网元测量来自于第三网元的定位参考信号，获得所述一个或多个测量结果，所述第三网元为接入网设备；
    以及，所述第一训练数据还包括所述第一网元的一个或多个位置信息。
  17. 根据权利要求4至16中任一项所述的方法，其特征在于，所述约束条件包括如下一项或多项：
    质量指标的门限和所述质量指标的判定准则；或，
    符合质量指标的判定准则的训练数据的数量门限和所述训练数据的数量的判定准则；或，
    单次有效性判定对应的训练数据收集的最大时长指示信息。
  18. 根据权利要求4至17中任一项所述的方法，所述第一信息指示如下一项或多项：
    质量指标的门限；
    质量指标的判定准则；
    符合质量指标的判定准则的训练数据的数量门限；
    符合质量指标的判定准则的训练数据的数量的判定准则；或
    单次有效性判定对应的候选训练数据收集的最大时长。
  19. 根据权利要求4至18中任一项所述的方法，所述约束条件基于所述AI模型的应用场景，所述AI模型的应用场景包括如下一项或多项：
    基于所述AI模型的CSI反馈或CSI预测、基于所述AI模型的定位，或，基于所述AI模型的波束管理。
  20. 一种AI模型训练中用于获取训练数据的方法，其特征在于，所述方法由第二网元或用于第二网元的芯片执行，所述方法包括：
    向第一网元发送第一信息，所述第一信息用于所述第一网元收集的所述AI模型的候选训练数据的有效性的判定，所述有效性的判定结果包括有效或无效；
    接收来自于所述第一网元的第二信息，所述第二信息指示所述有效性的判定结果。
  21. 根据权利要求20所述的方法，其特征在于，所述第二信息包括第一训练数据且所述第二信息指示所述第一网元收集的所述候选训练数据有效，所述第一训练数据为所述候选训练数据中的有效数据。
  22. 根据权利要求20所述的方法，其特征在于，所述第二信息指示所述第一网元收集的所述候选训练数据无效。
  23. 根据权利要求20至22中任一项所述的方法，其特征在于，所述第一信息用于所述第一网元收集的所述候选训练数据的有效性的判定的约束条件的确定。
  24. 根据权利要求23所述的方法，其特征在于，若所述候选训练数据中包含满足所述约束条件的 第一训练数据，所述候选训练数据有效；或者，
    若所述候选训练数据中不包含满足所述约束条件的第一训练数据，所述候选训练数据无效。
  25. 根据权利要求22至24中任一项所述的方法，其特征在于，在所述第二信息指示所述第一网元收集的所述候选训练数据无效的情况下，所述方法还包括：
    向所述第一网元发送第三信息，所述第三信息指示所述第一网元重新收集所述AI模型的候选训练数据。
  26. 根据权利要求25所述的方法，其特征在于，所述方法还包括：
    确定空口传输配置信息，所述空口传输配置信息对应更新的空口传输配置，所述空口传输配置信息指示所述第一网元基于所述更新的空口传输配置收集所述AI模型的候选训练数据；
    其中，所述更新的空口传输配置信息包括如下一项或多项的更新：
    参考信号的发送功率；
    参考信号使用的天线端口数；
    参考信号的频带宽度；
    参考信号的频域密度；或，
    参考信号的周期。
  27. 根据权利要求25或26所述的方法，其特征在于，所述第三信息还指示所述有效性的判定的最大次数k，k为正整数。
  28. 根据权利要求25或26所述的方法，其特征在于，所述第一信息还指示所述有效性的判定的最大次数k，k为正整数。
  29. 根据权利要求27或28所述的方法，其特征在于，所述方法还包括：
    接收来自于所述第一网元的第四信息，所述第四信息包括第二训练数据，且所述第四信息指示所述第一网元的第j次有效性判定的判定结果为有效，所述第二训练数据为所述第j次有效性的判定所针对的候选训练数据中的有效数据，j小于或等于k，j为正整数。
  30. 根据权利要求20至29中任一项所述的方法，其特征在于，所述第二网元为接入网设备或用于所述接入网设备的芯片，所述第一网元为终端设备或用于所述终端设备的芯片，所述方法还包括：
    向所述第一网元发送参考信号，所述参考信号用于所述第一网元获取对应于所述参考信号的一个或多个测量结果，所述AI模型的候选训练数据包括所述一个或多个测量结果。
  31. 根据权利要求30所述的方法，其特征在于，所述第一训练数据还包括所述一个或多个测量结果中的K个最优的测量结果对应的参考信号，K为大于或等于1的整数。
  32. 根据权利要求20至29中任一项所述的方法，其特征在于，所述第二网元为定位设备或用于所述定位设备的芯片，所述第一网元为接入网设备或用于所述接入网设备的芯片，所述AI模型的候选训练数据包括一个或多个测量结果和第三网元的位置信息，所述一个或多个测量结果是由所述第一网元测量所述第三网元发送的探测参考信号获得的。
  33. 根据权利要求20至29中任一项所述的方法，其特征在于，所述第二网元为定位设备或用于所述定位设备的芯片，所述第一网元为终端设备或用于所述终端设备的芯片，所述AI模型的候选训练数据包括一个或多个测量结果和所述第一网元的位置信息，所述一个或多个测量结果基于对所述第三网元发送的定位参考信号的测量，所述第三网元为接入网设备。
  34. 根据权利要求23-33中任一项所述的方法，其特征在于，所述约束条件包括如下一项或多项：
    质量指标的门限和所述质量指标的判定准则；或，
    符合质量指标的判定准则的训练数据的数量门限和所述训练数据的数量的判定准则；或，
    单次有效性判定对应的候选训练数据收集的最大时长。
  35. 根据权利要求20至34中任一项所述的方法，其特征在于，所述第一信息指示如下一项或多项：
    质量指标的门限；
    质量指标的判定准则；
    符合质量指标的判定准则的训练数据的数量门限；
    符合质量指标的判定准则的训练数据的数量的判定准则；或
    单次有效性判定对应的候选训练数据收集的最大时长。
  36. 根据权利要求23至35中任一项所述的方法，其特征在于，所述约束条件基于所述AI模型的应用场景，所述AI模型的应用场景包括如下一项或多项：
    基于所述AI模型的CSI反馈或CSI预测、基于所述AI模型的定位，或，基于所述AI模型的波束管理。
  37. 一种通信装置，其特征在于，包括用于实现如权利要求1-19中任一项所述的方法的模块，或者用于实现如权利要求20-36中任一项所述的方法的模块。
  38. 一种通信装置，其特征在于，包括：
    处理器，所述处理器和存储器耦合，所述处理器用于调用所述存储器存储的计算机程序指令，以执行如权利要求1-19任一项所述的方法，或者执行如权利要求20-36中任一项所述的方法。
  39. 一种通信装置，其特征在于，包括：处理器和通信接口，所述通信接口用于接收数据和/或信息，并将接收到的数据和/或信息传输至所述处理器；所述处理器处理所述数据和/或信息；以及，所述通信接口还用于输出经所述处理器处理之后的数据和/或信息，以使得所述通信装置执行如权利要求1-19中任一项所述的方法，或者，执行如权利要求20-36中任一项所述的方法。
  40. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有指令，当所述指令在计算机上运行时，使得计算机执行如权利要求1-19任一项所述的方法，或者执行如权利要求20-36中任一项所述的方法。
  41. 一种计算机程序产品，其特征在于，所述计算机可读存储介质上存储有指令，当所述指令在计算机上运行时，使得计算机执行如权利要求1-19任一项所述的方法，或者执行如权利要求20-36中任一项所述的方法。
  42. 一种通信系统，其特征在于，包括如权利要求37-39中任一项所述的通信装置。
PCT/CN2023/119343 2022-09-29 2023-09-18 Ai模型训练中用于获取训练数据的方法以及通信装置 WO2024067193A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211203052.XA CN117793767A (zh) 2022-09-29 2022-09-29 Ai模型训练中用于获取训练数据的方法以及通信装置
CN202211203052.X 2022-09-29

Publications (1)

Publication Number Publication Date
WO2024067193A1 true WO2024067193A1 (zh) 2024-04-04

Family

ID=90378656

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/119343 WO2024067193A1 (zh) 2022-09-29 2023-09-18 Ai模型训练中用于获取训练数据的方法以及通信装置

Country Status (2)

Country Link
CN (1) CN117793767A (zh)
WO (1) WO2024067193A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110475355A (zh) * 2018-05-11 2019-11-19 华为技术有限公司 一种波束训练的方法、装置及系统
WO2020164405A1 (zh) * 2019-02-15 2020-08-20 华为技术有限公司 一种定位方法和通信装置
WO2022011704A1 (zh) * 2020-07-17 2022-01-20 北京小米移动软件有限公司 定位测量数据上报方法、装置、终端及存储介质
US20220046386A1 (en) * 2020-08-04 2022-02-10 Qualcomm Incorporated Selective triggering of neural network functions for positioning of a user equipment
CN114788319A (zh) * 2019-11-22 2022-07-22 华为技术有限公司 个性化定制空口

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110475355A (zh) * 2018-05-11 2019-11-19 华为技术有限公司 一种波束训练的方法、装置及系统
WO2020164405A1 (zh) * 2019-02-15 2020-08-20 华为技术有限公司 一种定位方法和通信装置
CN114788319A (zh) * 2019-11-22 2022-07-22 华为技术有限公司 个性化定制空口
WO2022011704A1 (zh) * 2020-07-17 2022-01-20 北京小米移动软件有限公司 定位测量数据上报方法、装置、终端及存储介质
US20220046386A1 (en) * 2020-08-04 2022-02-10 Qualcomm Incorporated Selective triggering of neural network functions for positioning of a user equipment

Also Published As

Publication number Publication date
CN117793767A (zh) 2024-03-29

Similar Documents

Publication Publication Date Title
US11838787B2 (en) Functional architecture and interface for non-real-time ran intelligent controller
US20230010095A1 (en) Methods for cascade federated learning for telecommunications network performance and related apparatus
US20210345134A1 (en) Handling of machine learning to improve performance of a wireless communications network
EP3523901A1 (en) Forward error correction code selection in wireless systems
US20220294666A1 (en) Method for support of artificial intelligence or machine learning techniques for channel estimation and mobility enhancements
US20220287104A1 (en) Method and apparatus for support of machine learning or artificial intelligence techniques in communication systems
US20230217403A1 (en) Dynamic update of positioning reference signals based on user equipment location updates
US20220286927A1 (en) Method and apparatus for support of machine learning or artificial intelligence techniques for handover management in communication systems
US20220038931A1 (en) Radio link adaptation in wireless network
US20230006913A1 (en) Method and apparatus for channel environment classification
US20230269606A1 (en) Measurement configuration for local area machine learning radio resource management
CN111756457A (zh) 信道预测方法、装置及计算机存储介质
US20230198640A1 (en) Channel state information values-based estimation of reference signal received power values for wireless networks
US11647532B1 (en) Algorithm for mitigation of impact of uplink/downlink beam mismatch
WO2024067193A1 (zh) Ai模型训练中用于获取训练数据的方法以及通信装置
WO2021019287A1 (en) Methods and apparatuses for time-domain beam-sweeping
EP3987677A1 (en) Adaptive csi reporting and prb bundling in aas
CN113905385B (zh) 无线电资源参数配置
US20220361086A1 (en) Method and base station for determining transmission path in wireless communication system
US20230162006A1 (en) Server and agent for reporting of computational results during an iterative learning process
WO2023236143A1 (zh) 信息收发方法与装置
US20240162957A1 (en) Method and device for reporting csi based on ai model in wireless communication system
WO2024094038A1 (zh) 切换或更新ai模型的方法以及通信装置
WO2024002003A1 (zh) 信道反馈模型确定方法、终端及网络侧设备
US20240030977A1 (en) Partial channel state information reporting

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23870427

Country of ref document: EP

Kind code of ref document: A1