WO2024002003A1 - 信道反馈模型确定方法、终端及网络侧设备 - Google Patents

信道反馈模型确定方法、终端及网络侧设备 Download PDF

Info

Publication number
WO2024002003A1
WO2024002003A1 PCT/CN2023/102375 CN2023102375W WO2024002003A1 WO 2024002003 A1 WO2024002003 A1 WO 2024002003A1 CN 2023102375 W CN2023102375 W CN 2023102375W WO 2024002003 A1 WO2024002003 A1 WO 2024002003A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
channel
information
target
terminal
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2023/102375
Other languages
English (en)
French (fr)
Inventor
李刚
韩双锋
温子睿
石子烨
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Mobile Communications Group Co Ltd
Research Institute of China Mobile Communication Co Ltd
Original Assignee
China Mobile Communications Group Co Ltd
Research Institute of China Mobile Communication Co Ltd
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 China Mobile Communications Group Co Ltd, Research Institute of China Mobile Communication Co Ltd filed Critical China Mobile Communications Group Co Ltd
Publication of WO2024002003A1 publication Critical patent/WO2024002003A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • 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
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/391—Modelling the propagation channel
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0417—Feedback systems
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station

Definitions

  • the present disclosure relates to the field of communications, and in particular, to a method for determining a channel feedback model, a terminal, and a network side device.
  • the channel feedback model is generally divided into two parts: an encoder and a decoder.
  • an encoder needs to be deployed on the terminal side and a decoder needs to be deployed on the network side.
  • the terminal uses the encoder to compress the estimated channel state information (Channel State Information, CSI) into a series of bit streams.
  • CSI Channel State Information
  • the bit streams are fed back through the uplink
  • the channel is sent to the network side, and the network side finally recovers the original CSI based on the bit stream.
  • the training of the channel feedback model is deployed on the network side. Due to the need to adapt to different terminals, the network side needs to collect a large amount of channel data to train the channel feedback model, resulting in high overhead on the network side.
  • Embodiments of the present disclosure provide a method for determining a channel feedback model, a terminal, and a network side device to solve the problem that the network side needs to collect a large amount of channel data to train the channel feedback model, resulting in high overhead on the network side.
  • embodiments of the present disclosure provide a method for determining a channel feedback model, which is applied to a terminal.
  • the method includes:
  • the model information includes at least one of the following:
  • performing model training on the target channel feedback model based on the channel information includes:
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the weights of the encoding model in the target channel feedback model are maintained.
  • the coefficients and weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged;
  • the target layer neural network is at least one layer of neural network of the decoding model.
  • the method also includes:
  • the model training of the target channel feedback model based on the channel information includes:
  • Model training is performed on the target channel feedback model based on the training configuration.
  • the method before performing model training on the target channel feedback model based on the channel information, the method further includes:
  • a target channel feedback model is selected based on the channel information.
  • selecting a target channel feedback model based on the channel information includes:
  • target parameters include static characteristic parameters and/or dynamic environment parameters
  • a target channel feedback model is selected from at least two channel feedback models based on the target parameters.
  • embodiments of the present disclosure provide a method for determining a channel feedback model, which is applied to network side equipment.
  • the method includes:
  • Receive model information of the decoding model in the target channel feedback model sent by the terminal wherein the target channel feedback model is obtained by the terminal performing model training based on the channel information, or the target channel feedback model is the The terminal selects based on the channel information.
  • the model information includes at least one of the following:
  • the method before receiving the model information of the decoding model in the target channel feedback model sent by the terminal, the method further includes:
  • the training configuration is used to perform model training on the target channel feedback model.
  • the target channel feedback model is obtained by the network side device using training samples to train a basic model
  • the method Before sending the training configuration for model training to the terminal based on the channel information, the method further includes:
  • the first channel characteristics are compared with the second channel characteristics, and a training configuration for model training is determined based on the comparison results.
  • the second channel characteristics are average channel characteristics determined based on the training samples.
  • the first channel characteristics include at least one of the following:
  • an embodiment of the present disclosure provides a terminal, where the terminal includes:
  • a receiving module configured to receive the channel state information reference signal sent by the network side device, and perform channel measurement on the channel state information reference signal to obtain channel information;
  • a processing module configured to perform model training on a target channel feedback model based on the channel information, or select a target channel feedback model based on the channel information;
  • a first sending module configured to send the information in the target channel feedback model to the network side device. Decode model information for the model.
  • the model information includes at least one of the following:
  • processing module is specifically used to:
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the weights of the encoding model in the target channel feedback model are maintained.
  • the coefficients and weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged;
  • the target layer neural network is at least one layer of neural network of the decoding model.
  • the device also includes:
  • a second sending module configured to send the channel information to the network side device, and receive the training configuration for model training sent by the network side device based on the channel information
  • the processing module is specifically used for:
  • Model training is performed on the target channel feedback model based on the training configuration.
  • the device also includes:
  • a selection module configured to select a target channel feedback model based on the channel information.
  • processing module is specifically used to:
  • target parameters include static characteristic parameters and/or dynamic environment parameters
  • a target channel feedback model is selected from at least two channel feedback models based on the target parameters.
  • embodiments of the present disclosure provide a network side device, where the network side device includes:
  • the first sending module is configured to send a channel state information reference signal to the terminal, so that the terminal performs channel measurement on the channel state information reference signal to obtain channel information;
  • a receiving module configured to receive model information of a decoding model in a target channel feedback model sent by the terminal, where the target channel feedback model is obtained by the terminal through model training based on the channel information, or the target channel The feedback model is selected for the terminal based on the channel information.
  • the model information includes at least one of the following:
  • the device also includes:
  • a second sending module configured to receive the channel information sent by the terminal, and send training configuration for model training to the terminal based on the channel information
  • the training configuration is used to perform model training on the target channel feedback model.
  • the target channel feedback model is obtained by the network side device using training samples to train a basic model
  • the device also includes:
  • An acquisition module configured to acquire first channel characteristics based on the channel information
  • Determining module configured to compare the first channel characteristic with the second channel characteristic, and determine the training configuration for model training based on the comparison result, the second channel characteristic is an average value determined based on the training sample Channel characteristics.
  • the first channel characteristics include at least one of the following:
  • embodiments of the present disclosure provide a terminal, including a transceiver and a processor,
  • the transceiver is configured to receive a channel state information reference signal sent by a network side device, and perform channel measurement on the channel state information reference signal to obtain channel information;
  • the processor is configured to perform model training on a target channel feedback model based on the channel information, or select a target channel feedback model based on the channel information;
  • the transceiver is further configured to send model information of the decoding model in the target channel feedback model to the network side device.
  • the model information includes at least one of the following:
  • the processor is used for:
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the target channel is maintained
  • the weight coefficients of the encoding model in the feedback model and the weight coefficients of other layer neural networks in the decoding model except the target layer neural network remain unchanged;
  • the target layer neural network is at least one layer of neural network of the decoding model.
  • the transceiver is further configured to: send the channel information to the network side device, and receive a training configuration for model training sent by the network side device based on the channel information;
  • the processor is also used to:
  • Model training is performed on the target channel feedback model based on the training configuration.
  • the processor is further configured to select a target channel feedback model based on the channel information.
  • the processor is also used to:
  • target parameters include static characteristic parameters and/or dynamic environment parameters
  • a target channel feedback model is selected from at least two channel feedback models based on the target parameters.
  • embodiments of the present disclosure provide a network side device, including a transceiver and a processor,
  • the transceiver is configured to send a channel state information reference signal to a terminal, so that the terminal performs channel measurement on the channel state information reference signal to obtain channel information;
  • the transceiver is further configured to receive model information of the decoding model in the target channel feedback model sent by the terminal, wherein the target channel feedback model is obtained by the terminal through model training based on the channel information, or the The target channel feedback model is selected by the terminal based on the channel information.
  • the model information includes at least one of the following:
  • the transceiver is also used for:
  • the training configuration is used to perform model training on the target channel feedback model.
  • the target channel feedback model is obtained by the network side device using training samples to train a basic model
  • the processor is used for:
  • the first channel characteristics are compared with the second channel characteristics, and a training configuration for model training is determined based on the comparison results.
  • the second channel characteristics are average channel characteristics determined based on the training samples.
  • the first channel characteristics include at least one of the following:
  • an embodiment of the present disclosure provides a terminal, including: a processor, a memory, and a program stored on the memory and executable on the processor.
  • the program is executed by the processor, the above is implemented. The steps of the method for determining the channel feedback model described in the first aspect.
  • an embodiment of the present disclosure provides a network-side device, including: a processor, a memory, and a program stored on the memory and executable on the processor.
  • a program stored on the memory and executable on the processor.
  • embodiments of the present disclosure provide a computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, the channel feedback model described in the first aspect is implemented.
  • the steps of the determination method; or when the computer program is executed by the processor, the steps of the channel feedback model determination method described in the second aspect are implemented.
  • the terminal receives the channel state information reference signal sent by the network side device, performs channel measurement on the channel state information reference signal to obtain channel information; performs model training on the target channel feedback model based on the channel information, or Select a target channel feedback model based on the channel information; and send model information of the decoding model in the target channel feedback model to the network side device.
  • the terminal performs model training on the target channel feedback model or selecting the target channel feedback model, the overhead on the network side can be reduced.
  • Figure 1 is one of the flow charts of a method for determining a channel feedback model provided by an embodiment of the present disclosure
  • Figure 2 is one of the structural schematic diagrams of a channel feedback model provided by an embodiment of the present disclosure
  • Figure 3 is the second flow chart of a method for determining a channel feedback model provided by an embodiment of the present disclosure
  • Figure 4 is a schematic structural diagram of model information provided by an embodiment of the present disclosure.
  • Figure 5 is a schematic diagram of a model training result provided by an embodiment of the present disclosure.
  • Figure 6 is a schematic classification diagram of a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 7a is one of the schematic diagrams of a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 7b is a second schematic diagram of a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 7c is a third schematic diagram of a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 7d is a fourth schematic diagram of a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 7e is a fifth schematic diagram of a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 8 is the third flowchart of a method for determining a channel feedback model provided by an embodiment of the present disclosure
  • Figure 9 is the fourth flowchart of a method for determining a channel feedback model provided by an embodiment of the present disclosure.
  • Figure 10 is the fifth flowchart of a method for determining a channel feedback model provided by an embodiment of the present disclosure
  • Figure 11 is one of the structural schematic diagrams of a terminal provided by an embodiment of the present disclosure.
  • Figure 12 is one of the structural schematic diagrams of a network side device provided by an embodiment of the present disclosure.
  • Figure 13 is a second structural schematic diagram of a terminal provided by an embodiment of the present disclosure.
  • Figure 14 is a second structural schematic diagram of a network side device provided by an embodiment of the present disclosure.
  • a method for determining a channel feedback model, a terminal and a network side device are proposed to solve the problem that the network side needs to collect a large amount of channel data to train the channel feedback model, resulting in high overhead on the network side.
  • Figure 1 is a flow chart of a method for determining a channel feedback model provided by an embodiment of the present disclosure.
  • Process diagram, used for terminals, as shown in Figure 1, the method includes the following steps:
  • Step 101 Receive the channel state information reference signal sent by the network side device, and perform channel measurement on the channel state information reference signal to obtain channel information.
  • the channel state information reference signal (CSI-RS) is sent by the network side device and is used by the terminal to perform channel measurements and obtain channel information.
  • the channel information may be channel state information (Channel State Information, CSI).
  • Step 102 Perform model training on a target channel feedback model based on the channel information, or select a target channel feedback model based on the channel information.
  • the target channel feedback model may include a coding model and a decoding model.
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the target channel feedback model is maintained The weight coefficients of the encoding model and the weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged.
  • the target layer neural network is at least one layer of neural networks of the decoding model. .
  • the target channel feedback model can be trained in the following two ways:
  • the terminal determines the training configuration for online training.
  • the terminal can determine the optimal training configuration and the number of model layers and weight coefficients that need to be fed back to the network side device.
  • the terminal can start training from the last layer of the decoding model until there is no significant improvement in performance.
  • the terminal can train the penultimate layer according to priority 1, train the penultimate layer and penultimate layer according to priority 2, and train the penultimate layer, penultimate layer and penultimate layer according to priority 3.
  • ... train the penultimate layer, the penultimate layer, the penultimate layer, ..., the penultimate m layer according to the priority m, and stop training until the training accuracy does not increase.
  • priority 1 is the highest priority.
  • the network side device determines the training configuration for online training.
  • Network-side devices can configure optimal training configurations for terminals based on historical big data analysis results.
  • the model structure can be divided into a frozen layer and a fine-tuning layer.
  • the terminal does not update the weight coefficients during back propagation during model training.
  • the terminal performs gradient descent during back propagation during model training. weight system Number of updates.
  • the terminal only needs to feedback the weight coefficients of the fine-tuning layer, which can reduce training overhead and feedback overhead.
  • the training configuration is as follows: fine-tuning layer: 1, 2...m; freezing layer: m...n; number of training iterations.
  • Step 103 Send model information of the decoding model in the target channel feedback model to the network side device.
  • the terminal may send model information of the decoding model in the target channel feedback model after model training to the network side device.
  • the terminal when the terminal uses the channel information obtained by itself to perform online training on the target channel feedback model, it can obtain the uplink transmission resources through the uplink service resource transmission request, and then in the service data adaptation protocol (The model information of the decoding model in the target channel feedback model is transmitted in the Service Data Adaptation Protocol (SDAP) data packet.
  • SDAP Service Data Adaptation Protocol
  • the model information can include: model identifier (Identifier, ID); model layer that needs to be updated. numerical information (for example, L1 to Ln); model weight coefficients that need to be updated.
  • the uplink service resource transmission request may include: the terminal sends Scheduling Request (SR) signaling to the Radio Access Network (Radio Access Network, RAN) through the Physical Uplink Control Channel (PUCCH), and the RAN
  • SR Scheduling Request
  • the physical downlink control channel Physical downlink control channel, PDCCH
  • DCI Downlink Control Information
  • PDCCH Physical downlink control channel
  • DCI Downlink Control Information
  • UL uplink
  • MAC Media Access Control
  • the MAC control unit Control Element, CE
  • BSR Buffer Status Report
  • the network side equipment (such as the base station) can periodically send CSI-RS; after receiving the CSI-RS, the terminal performs channel estimation, uses the coding model in the target channel feedback model to perform compression and quantization, and then sends the CSI-RS to the network.
  • the side device sends the compressed and quantized channel information; the network side device receives the compressed and quantized channel information, uses the updated decoding model to decompress the channel information, and obtains complete downlink channel estimation information.
  • the encoding model and decoding model can be trained uniformly in an end-to-end manner during model training, and can also be used in pairs during deployment; if mismatched encoding models and decoding models are used, it will cause a significant decrease in channel information recovery accuracy.
  • model training is performed by a network-side device. Yes, since model training requires a lot of computing power, network-side equipment needs to collect a large amount of channel data, which is expensive. Moreover, limited by the collection scenarios and collection scale of training data, and due to the changeable channel environment where the terminal is located, it is difficult for a fixed channel feedback model to achieve good performance in all wireless scenarios, that is, there is a general problem chemical issues.
  • the terminal performs model training on the target channel feedback model, or selects the target channel feedback model based on channel information, which can solve the problem of reduced recovery accuracy caused by the generalization of the encoding model and the decoding model in different environments. question.
  • the terminal receives the channel state information reference signal sent by the network side device, performs channel measurement on the channel state information reference signal to obtain channel information; performs model training on the target channel feedback model based on the channel information, or Select a target channel feedback model based on the channel information; and send model information of the decoding model in the target channel feedback model to the network side device.
  • the terminal performs model training on the target channel feedback model or selecting the target channel feedback model, the overhead on the network side can be reduced.
  • the model information includes at least one of the following:
  • the terminal sends the model identifier of the decoding model to the network side device, and/or all the weight coefficients of the decoding model or part of the weight coefficients of the decoding model, so that the network side device can based on the model identifier sent by the terminal and /or the weight coefficient updates the decoding model.
  • performing model training on the target channel feedback model based on the channel information includes:
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the weights of the encoding model in the target channel feedback model are maintained.
  • the coefficients and weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged;
  • the target layer neural network is at least one layer of neural network of the decoding model.
  • the target channel feedback model can be obtained by the network side device using training samples to train a basic model.
  • the encoding model in the basic model is composed of four fully connected layers
  • the decoding model in the basic model is also composed of four fully connected layers.
  • the initial weight coefficient of the basic model can be obtained by training the network side device through the collected big data.
  • the terminal can choose three types.
  • the base model is retrained on channel data sets of different sizes. For example, as shown in Figure 5, the basic model can be retrained through large data sets, medium data sets, and small data sets. Moreover, different layers of the base model can be frozen to retrain the base model. As can be seen from Figure 5, there is a better training strategy for retraining on the terminal side.
  • the Normalized Mean Square Error (NMSE) of the last layer of training is higher, but the NMSE is fast after training the last two layers. decline, the subsequent marginal effects will be diminishing, and retraining the last two layers is a better choice to balance efficiency and performance.
  • the performance of the decoding model can be optimized by optimizing the weight coefficients of some layers of the decoding model. For the terminal, only feeding back part of the layer weight coefficients of the decoding model can greatly reduce the overhead of air interface transmission.
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the weight coefficients of the target layer neural network in the target channel feedback model are maintained.
  • the weight coefficients of the encoding model and the weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged, so that only the weight coefficients of some layers are trained during model training, and then the network side
  • the device only transmits the weight coefficients of some layers of the decoding model, which can reduce the overhead of air interface transmission.
  • the method also includes:
  • the model training of the target channel feedback model based on the channel information includes:
  • Model training is performed on the target channel feedback model based on the training configuration.
  • the terminal can send an uplink sounding reference signal (SRS) to the network side device, and receive a training configuration for model training sent by the network side device.
  • the training configuration can be based on channel characteristics. Determined, the channel characteristics may be determined based on the channel information sent by the terminal and the SRS.
  • Wireless channel characteristics include the number of uplink and downlink multipaths, angles, multipath delays and other parameters.
  • TDD Time Division Duplex
  • wireless channels have good mutuality, and terminals can communicate with the network side.
  • the device sends the uplink sounding reference signal SRS for channel information acquisition, and the network side device measures the SRS.
  • SRS uplink sounding reference signal
  • the device sends the uplink sounding reference signal SRS for channel information acquisition, and the network side device measures the SRS.
  • Obtain the uplink channel matrix H and use the uplink and downlink reciprocity to obtain the downlink channel matrix H', and calculate the precoding matrix and beamforming vector for downlink data transmission through the downlink channel matrix H'.
  • the uplink and downlink are deployed at different frequencies, and there is no complete channel reciprocity.
  • the angle and delay of each path on the wireless channel are mainly affected by the spatial angle of the transceiver. It is determined by factors such as the relationship, the position and material of the reflector in the wireless environment, and there is no strong relationship with the frequency of the wireless signal.
  • the frequency of the wireless channel will greatly affect the path loss, penetration loss, polarization leakage factor and other components of the wireless channel, which will have a greater impact on the amplitude changes and phase changes experienced by each path.
  • the network side device can obtain partial channel characteristics of the downlink channel based on the partial reciprocity of the channel, and can also obtain the channel characteristics based on the channel information fed back by the terminal.
  • the terminal sends the channel information to the network side device, and receives the training configuration for model training sent by the network side device based on the channel information; the terminal feeds back the target channel based on the training configuration
  • the model performs model training so that the training configuration can be determined through the network side device.
  • the method before performing model training on the target channel feedback model based on the channel information, the method further includes:
  • a target channel feedback model is selected based on the channel information.
  • selecting a target channel feedback model based on the channel information includes:
  • target parameters include static characteristic parameters and/or dynamic environment parameters
  • a target channel feedback model is selected from at least two channel feedback models based on the target parameters.
  • the static characteristic parameters may include but are not limited to: network configuration parameters, such as macro stations/small stations, the number of multi-antenna ports on the network side, etc.; terminal capability parameters, such as the number of terminal multi-antenna ports, terminal Artificial Intelligence (AI) Reasoning ability, etc.
  • the above static characteristic parameters generally belong to the information that needs to be reported when the terminal initially accesses the network, and remains basically unchanged during the occurrence of communication services.
  • Dynamic environment parameters can correspond to various dynamic scenarios. Dynamic environment parameters can include channel environment parameters and channel quality parameters, etc. Channel environment parameters can be channel environment parameters under line-of-sight or non-line-of-sight transmission. Number, channel quality parameters may include signal-to-noise ratio range, etc.
  • the network side device can form different training sets based on target parameters during the training stage of the basic model, train multiple models on the basis of different training sets, and divide the multiple models into different applicable scenarios. Multiple categories.
  • the terminal can select a target channel feedback model from multiple models.
  • the multiple models can be first classified according to static feature parameters to obtain multiple semi-static categories, and then determined by the static feature parameters.
  • Each subcategory continues to be classified according to dynamic environment parameters, and each subcategory can be divided into multiple dynamic subcategory models.
  • encoding models or decoding models can be shared between multiple models within each subclass.
  • the encoding model can be called an encoder, and the decoding model can be called a decoder.
  • the encoding model can be called an encoder, and the decoding model can be called a decoder.
  • five encoding model-decoding model modes are listed. Multiple encoding models can share one decoding model, or multiple decoding models can share one encoding model. The five modes listed can be divided into two categories according to whether there are multiple encoding models or decoding models to choose from, namely acde/b or abde/c.
  • Each encoding model-decoding model pair corresponds to a set of dynamic categories. In actual use, a specific group of encoding models-decoding models can be selected to compress and restore CSI.
  • selecting a target channel feedback model from at least two channel feedback models based on the target parameters may be to mark applicable static feature parameters and/or dynamic environment parameters for each pair of encoding model-decoding model, according to The target parameters select the corresponding encoding model-decoding model by looking up the dictionary.
  • At least two channel feedback models can be divided into multiple subcategories according to applicable static characteristic parameters and/or dynamic environment parameters, and a target channel feedback model is selected from the at least two channel feedback models based on the target parameters, It can be that the encoding model and decoding model of the corresponding subcategory are selected according to the target parameters by searching a dictionary, and the encoding model and decoding model in the selected subcategory are used in turn to compress and restore the channel information, and the recovery accuracy is calculated, and the subcategories are traversed. The models in the model are sorted according to the recovery accuracy, and the model with the highest recovery accuracy is selected as the target channel feedback model.
  • the terminal can notify the network side device of the decoding model of the selected target channel feedback model.
  • the notification method can be by sending signaling carrying the decoding model to the network side device.
  • the transmission of model information of the decoding model can be implemented by extending Radio Resource Control (RRC) or MAC CE signaling.
  • RRC Radio Resource Control
  • MAC CE MAC CE signaling
  • a Logic Channel Group (LCG) ID field can be added to the MAC CE signaling, and this field carries the model identification of the decoding model.
  • selecting a target channel feedback model from at least two channel feedback models based on static characteristic parameters and/or dynamic environment parameters can obtain a channel feedback model that matches the current static characteristic parameters and/or dynamic environment parameters.
  • Figure 8 is a flow chart of a method for determining a channel feedback model provided by an embodiment of the present disclosure, which is used for network side equipment. As shown in Figure 8, the method includes the following steps:
  • Step 201 Send a channel state information reference signal to the terminal, so that the terminal performs channel measurement on the channel state information reference signal to obtain channel information.
  • Step 202 Receive model information of the decoding model in the target channel feedback model sent by the terminal, where the target channel feedback model is obtained by the terminal performing model training based on the channel information, or the target channel feedback model Select for the terminal based on the channel information.
  • this embodiment is an implementation of the network side device corresponding to the embodiment shown in Figure 1.
  • the network side device corresponding to the embodiment shown in Figure 1.
  • no further details will be given in this embodiment.
  • the model information includes at least one of the following:
  • the method before receiving the model information of the decoding model in the target channel feedback model sent by the terminal, the method further includes:
  • the training configuration is used to perform model training on the target channel feedback model.
  • the target channel feedback model is obtained by the network side device using training samples to train a basic model
  • the method Before sending the training configuration for model training to the terminal based on the channel information, the method further includes:
  • the first channel characteristics are compared with the second channel characteristics, and a training configuration for model training is determined based on the comparison results.
  • the second channel characteristics are average channel characteristics determined based on the training samples.
  • the first channel characteristics are obtained based on the channel information; the first channel characteristics are compared with the second channel characteristics, and the training configuration for model training is determined based on the comparison results, and the second channel characteristics are compared with each other.
  • the channel characteristics are average channel characteristics determined based on the training samples.
  • the training configuration can be determined by comparing the average channel characteristics determined by the training samples of the training basic model and the first channel characteristics, and appropriate training configurations can be determined for different first channel characteristics, thereby achieving better model training effects. .
  • the first channel characteristics include at least one of the following:
  • the angle domain information reflects the sparsity of the beam/antenna.
  • different training configurations can be determined by whether the proportion of the first three beams with the largest path gain in the overall power gain exceeds the first threshold indicated by the angle domain information.
  • the first threshold is determined by the second channel characteristics.
  • the proportion of the first three beams with the largest path gains indicated by the angle domain information in the second channel characteristics in the overall power gain is the first threshold.
  • the first threshold can be 80%.
  • the delay domain information reflects the sparsity of the delay domain/multipath, non-line of sight (NLOS) or line of sight (Line of Sight, LOS) channels.
  • the delay domain information can be used Whether the indicated K factor (that is, the power ratio occupied by the main path) exceeds the second threshold determines different training configurations.
  • the second threshold is determined by the second channel characteristics.
  • the delay domain information in the second channel characteristics indicates The K factor of is the second threshold.
  • the second threshold may be 50%.
  • the Doppler domain information reflects the sparsity in the time domain. For example, Different training configurations are determined by whether the frequency offset power proportion of the first five main paths indicated by the Doppler domain information exceeds a third threshold.
  • the third threshold is determined by the second channel characteristics. In one embodiment, the third channel characteristics
  • the frequency offset power proportion of the first five main paths indicated by the mid-Doppler domain information is the third threshold.
  • the third threshold may be 60%.
  • the proportion of the beams with the largest gain in the first three paths indicated by the angle domain information in the overall power gain is A
  • the K factor indicated by the delay domain information is B
  • the K factor indicated by the Doppler domain information is B.
  • the frequency offset power proportion of the first five main paths is C.
  • the encoding model consists of four fully connected layers
  • the decoding model consists of four fully connected layers.
  • the model structure is divided into a frozen layer and a fine-tuning layer.
  • the terminal does not update the weight coefficients during model training and backpropagation.
  • the fine-tuning layer the terminal performs weight coefficients through gradient descent during model training and backpropagation. Coefficient update.
  • the training configuration for model training can be: the number of frozen layers is set to be larger, and the fine-tuning layer is the fourth layer of the decoding model;
  • the training configuration for model training can be: the number of frozen layers is set to a small number, and the fine-tuning layers are the third and fourth layers of the decoding model;
  • the training configuration for model training can be: a small number of consecutive samples, a sample number of 100, and a duration of 500ms;
  • the training configuration for model training can be: a large number of consecutive samples, a sample number of 1000, and a duration of 1s.
  • the terminal conducts online training on the encoding model and the decoding model to determine the channel feedback model. Specifically, as shown in Figure 9, it includes the following process:
  • the channel feedback model includes a coding model and a decoding model.
  • the base station can send the channel feedback model to the terminal, or the channel feedback model can be pre-configured in the terminal;
  • the base station sends the CSI-RS configuration through the RRC reconfiguration message
  • the base station periodically sends CSI-RS to the terminal according to the CSI-RS configuration
  • the terminal receives CSI-RS and measures CSI-RS to obtain channel state information
  • the terminal sends a CSI report to the base station and sends an SRS to the base station;
  • the base station selects the configuration for online training based on channel characteristics
  • the base station sends the online training configuration to the terminal
  • the terminal performs online training on the encoding model and decoding model based on channel state information
  • the terminal sends the model information of the trained decoding model to the base station;
  • the base station updates the decoding model based on the received model information
  • the base station periodically sends CSI-RS to the terminal;
  • the terminal uses the trained coding model to compress CSI and performs quantization processing on the compressed CSI;
  • the terminal reports the compressed and quantized CSI to the base station
  • the base station uses the updated decoding model to decode CSI.
  • the channel feedback model is determined by the terminal selecting the channel feedback model. Specifically, as shown in Figure 10, it includes the following process:
  • the channel feedback model includes a coding model and a decoding model.
  • the base station can send the channel feedback model to the terminal, or the channel feedback model can be pre-configured in the terminal;
  • the base station sends the CSI-RS configuration through the RRC reconfiguration message
  • the base station periodically sends CSI-RS to the terminal according to the CSI-RS configuration
  • the terminal receives CSI-RS and measures CSI-RS to obtain channel state information
  • the terminal selects a decoding model based on the channel state information obtained by measurement
  • the terminal sends the model information of the selected decoding model to the base station;
  • the base station updates the decoding model based on the received model information
  • the base station periodically sends CSI-RS to the terminal;
  • the terminal uses the coding model to compress CSI and performs quantization processing on the compressed CSI;
  • the terminal reports the compressed and quantized CSI to the base station
  • the base station uses the updated decoding model to decode CSI.
  • Figure 11 is a schematic structural diagram of a terminal provided by an embodiment of the present disclosure. As shown in 11, the terminal 300 includes:
  • the receiving module 301 is configured to receive the channel state information reference signal sent by the network side device, and perform channel measurement on the channel state information reference signal to obtain channel information;
  • the processing module 302 is configured to perform model training on a target channel feedback model based on the channel information, or select a target channel feedback model based on the channel information;
  • the first sending module 303 is configured to send model information of the decoding model in the target channel feedback model to the network side device.
  • the model information includes at least one of the following:
  • processing module is specifically used to:
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the weights of the encoding model in the target channel feedback model are maintained.
  • the coefficients and weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged;
  • the target layer neural network is at least one layer of neural network of the decoding model.
  • the device also includes:
  • a second sending module configured to send the channel information to the network side device, and receive the training configuration for model training sent by the network side device based on the channel information
  • the processing module is specifically used for:
  • Model training is performed on the target channel feedback model based on the training configuration.
  • the device also includes:
  • a selection module configured to select a target channel feedback model based on the channel information.
  • processing module is specifically used to:
  • target parameters include static characteristic parameters and/or dynamic environment parameters
  • a target channel feedback model is selected from at least two channel feedback models based on the target parameters.
  • the terminal can implement each process implemented by the method embodiment shown in Figure 1 and can achieve the same beneficial effects. To avoid repetition, details will not be described here.
  • Figure 12 is a schematic structural diagram of a network side device provided by an embodiment of the present disclosure.
  • the network side device 400 includes:
  • the first sending module 401 is configured to send a channel state information reference signal to the terminal, so that the terminal performs channel measurement on the channel state information reference signal to obtain channel information;
  • the receiving module 402 is configured to receive the model information of the decoding model in the target channel feedback model sent by the terminal, wherein the target channel feedback model is obtained by the terminal through model training based on the channel information, or the target The channel feedback model is selected by the terminal based on the channel information.
  • the model information includes at least one of the following:
  • the device also includes:
  • a second sending module configured to receive the channel information sent by the terminal, and send training configuration for model training to the terminal based on the channel information
  • the training configuration is used to perform model training on the target channel feedback model.
  • the target channel feedback model is obtained by the network side device using training samples to train a basic model
  • the device also includes:
  • An acquisition module configured to acquire first channel characteristics based on the channel information
  • Determining module configured to compare the first channel characteristic with the second channel characteristic, and determine the training configuration for model training based on the comparison result, the second channel characteristic is an average value determined based on the training sample Channel characteristics.
  • the first channel characteristics include at least one of the following:
  • the network side device can implement each process implemented by the method embodiment shown in Figure 8 and can achieve the same beneficial effects. To avoid duplication, details will not be described here.
  • An embodiment of the present disclosure also provides a terminal, including: a processor, a memory and a device stored in the A program on the memory that can be run on the processor.
  • a terminal including: a processor, a memory and a device stored in the A program on the memory that can be run on the processor.
  • an embodiment of the present disclosure also provides a terminal, including a bus 501, a transceiver 502, an antenna 503, a bus interface 504, a processor 505 and a memory 506.
  • the transceiver 502 is used to receive a channel state information reference signal sent by a network side device, and perform channel measurement on the channel state information reference signal to obtain channel information;
  • the processor 505 is configured to perform model training on a target channel feedback model based on the channel information, or select a target channel feedback model based on the channel information;
  • the transceiver 502 is also configured to send model information of the decoding model in the target channel feedback model to the network side device.
  • the model information includes at least one of the following:
  • the processor 505 is used for:
  • the weight coefficients of the target layer neural network in the decoding model are trained based on the channel information, and the weights of the encoding model in the target channel feedback model are maintained.
  • the coefficients and weight coefficients of other layers of neural networks in the decoding model except the target layer neural network remain unchanged;
  • the target layer neural network is at least one layer of neural network of the decoding model.
  • the transceiver 502 is also configured to: send the channel information to the network side device, and receive a training configuration for model training sent by the network side device based on the channel information;
  • the processor 505 is also used to:
  • Model training is performed on the target channel feedback model based on the training configuration.
  • the processor 505 is also configured to select a target channel feedback model based on the channel information.
  • processor 505 is also used to:
  • target parameters include static characteristic parameters and/or dynamic environment parameters
  • a target channel feedback model is selected from at least two channel feedback models based on the target parameters.
  • bus 501 may include any number of interconnected buses and bridges, bus 501 will include one or more processors represented by processor 505 and memory represented by memory 506 various circuits linked together. Bus 501 may also link together various other circuits such as peripherals, voltage regulators, power management circuits, etc., which are all well known in the art and therefore will not be described further herein.
  • Bus interface 504 provides an interface between bus 501 and transceiver 502.
  • Transceiver 502 may be one element or may be multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium.
  • the data processed by the processor 505 is transmitted on the wireless medium through the antenna 503. Further, the antenna 503 also receives the data and transmits the data to the processor 505.
  • Processor 505 is responsible for managing bus 501 and general processing, and may also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions.
  • Memory 506 may be used to store data used by processor 505 when performing operations.
  • the processor 505 can be a central processing unit (Central Processing Unit, CPU), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable logic gate array (Field Programmable Gate Array, FPGA) or a complex programmable logic gate array.
  • Logic device Complex Programmable logic device, CPLD).
  • An embodiment of the present disclosure also provides a network-side device, including: a processor, a memory, and a program stored on the memory and executable on the processor.
  • a network-side device including: a processor, a memory, and a program stored on the memory and executable on the processor.
  • this embodiment of the present disclosure also provides a network side device, including a bus 601, a transceiver 602, an antenna 603, a bus interface 604, a processor 605 and a memory 606.
  • a network side device including a bus 601, a transceiver 602, an antenna 603, a bus interface 604, a processor 605 and a memory 606.
  • the transceiver 602 is used to send a channel state information reference signal to the terminal, so that the terminal performs channel measurement on the channel state information reference signal to obtain channel information;
  • the transceiver 602 is also configured to receive model information of the decoding model in the target channel feedback model sent by the terminal, wherein the target channel feedback model is obtained by the terminal through model training based on the channel information, or The target channel feedback model is selected by the terminal based on the channel information.
  • the model information includes at least one of the following:
  • the transceiver 602 is also used to:
  • the training configuration is used to perform model training on the target channel feedback model.
  • the target channel feedback model is obtained by the network side device using training samples to train a basic model
  • the processor 605 is used for:
  • the first channel characteristics are compared with the second channel characteristics, and a training configuration for model training is determined based on the comparison results.
  • the second channel characteristics are average channel characteristics determined based on the training samples.
  • the first channel characteristics include at least one of the following:
  • bus 601 may include any number of interconnected buses and bridges, bus 601 will include one or more processors 605 represented by processor 605 and memory 606 The various circuits of memory are linked together. Bus 601 may also link together various other circuits such as peripherals, voltage regulators, power management circuits, etc., which are all well known in the art and therefore will not be described further herein.
  • Bus interface 604 provides an interface between bus 601 and transceiver 602.
  • Transceiver 602 may be one element or may be multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium.
  • the data processed by the processor 605 is transmitted on the wireless medium through the antenna 603. Further, the antenna 603 also receives the data and transmits the data to the processor 605.
  • Processor 605 is responsible for managing bus 601 and general processing, and may also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. And memory 606 May be used to store data used by processor 605 when performing operations.
  • the processor 605 can be a CPU, ASIC, FPGA or CPLD.
  • Embodiments of the present disclosure also provide a computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, each process of the above-mentioned channel feedback model determination method embodiment is implemented, and the same can be achieved. The technical effects will not be repeated here to avoid repetition.
  • the computer-readable storage medium is such as read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk, etc.
  • the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk). ), includes several instructions to cause a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present disclosure.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Electromagnetism (AREA)
  • Quality & Reliability (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本公开提供一种信道反馈模型确定方法、终端及网络侧设备，涉及通信技术领域，其中，所述方法包括：接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。

Description

信道反馈模型确定方法、终端及网络侧设备
相关申请的交叉引用
本申请主张在2022年6月28日在中国提交的中国专利申请号No.202210748510.1的优先权，其全部内容通过引用包含于此。
技术领域
本公开涉及通信领域，尤其涉及一种信道反馈模型确定方法、终端及网络侧设备。
背景技术
目前，基于机器学习的信道状态信息压缩反馈受到了广泛关注。对于基于机器学习的信道状态信息压缩反馈而言，信道反馈模型一般分为编码器(encoder)及解码器(decoder)两部分。在实际部署时，需要在终端侧部署编码器，在网络侧部署解码器，终端使用编码器将估计得到的信道状态信息(Channel State Information，CSI)压缩为一串比特流，比特流经上行反馈信道发送至网络侧，网络侧最终基于该比特流恢复出原始的CSI。目前，信道反馈模型的训练部署在网络侧，由于需要适应不同的终端，网络侧需要采集大量的信道数据对信道反馈模型进行训练，导致网络侧的开销较大。
发明内容
本公开实施例提供一种信道反馈模型确定方法、终端及网络侧设备，以解决网络侧需要采集大量的信道数据对信道反馈模型进行训练，导致网络侧的开销较大的问题。
为解决上述技术问题，本公开是这样实现的：
第一方面，本公开实施例提供了一种信道反馈模型确定方法，应用于终端，所述方法包括：
接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述基于所述信道信息对目标信道反馈模型进行模型训练，包括：
在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
可选地，所述方法还包括：
向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
所述基于所述信道信息对目标信道反馈模型进行模型训练，包括：
基于所述训练配置对目标信道反馈模型进行模型训练。
可选地，所述基于所述信道信息对目标信道反馈模型进行模型训练之前，所述方法还包括：
基于所述信道信息选择目标信道反馈模型。
可选地，所述基于所述信道信息选择目标信道反馈模型，包括：
基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或动态环境参数；
基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
第二方面，本公开实施例提供了一种信道反馈模型确定方法，应用于网络侧设备，所述方法包括：
向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述接收所述终端发送的目标信道反馈模型中的解码模型的模型信息之前，所述方法还包括:
接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
可选地，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
所述基于所述信道信息向所述终端发送用于模型训练的训练配置之前，所述方法还包括：
基于所述信道信息获取第一信道特征；
将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
可选地，所述第一信道特征包括如下至少一项：
角度域信息；
时延域信息；
多普勒域信息。
第三方面，本公开实施例提供了一种终端，所述终端包括：
接收模块，用于接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
处理模块，用于基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
第一发送模块，用于向所述网络侧设备发送所述目标信道反馈模型中的 解码模型的模型信息。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述处理模块具体用于：
在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
可选地，所述装置还包括：
第二发送模块，用于向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
所述处理模块具体用于：
基于所述训练配置对目标信道反馈模型进行模型训练。
可选地，所述装置还包括：
选择模块，用于基于所述信道信息选择目标信道反馈模型。
可选地，所述处理模块具体用于：
基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或动态环境参数；
基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
第四方面，本公开实施例提供了一种网络侧设备，所述网络侧设备包括：
第一发送模块，用于向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
接收模块，用于接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述装置还包括:
第二发送模块，用于接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
可选地，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
所述装置还包括：
获取模块，用于基于所述信道信息获取第一信道特征；
确定模块，用于将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
可选地，所述第一信道特征包括如下至少一项：
角度域信息；
时延域信息；
多普勒域信息。
第五方面，本公开实施例提供了一种终端，包括收发机和处理器，
所述收发机，用于接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
所述处理器，用于基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
所述收发机，还用于向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述处理器用于：
在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道 反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
可选地，所述收发机还用于：向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
所述处理器还用于：
基于所述训练配置对目标信道反馈模型进行模型训练。
可选地，所述处理器还用于：基于所述信道信息选择目标信道反馈模型。
可选地，所述处理器还用于：
基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或动态环境参数；
基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
第六方面，本公开实施例提供了一种网络侧设备，包括收发机和处理器，
所述收发机，用于向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
所述收发机，还用于接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述收发机，还用于：
接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
可选地，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
所述处理器用于：
基于所述信道信息获取第一信道特征；
将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
可选地，所述第一信道特征包括如下至少一项：
角度域信息；
时延域信息；
多普勒域信息。
第七方面，本公开实施例提供一种终端，包括：处理器、存储器及存储在所述存储器上并可在所述处理器上运行的程序，所述程序被所述处理器执行时实现上述第一方面所述的信道反馈模型确定方法的步骤。
第八方面，本公开实施例提供一种网络侧设备，包括：处理器、存储器及存储在所述存储器上并可在所述处理器上运行的程序，所述程序被所述处理器执行时实现上述第二方面所述的信道反馈模型确定方法的步骤。
第九方面，本公开实施例提供一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现上述第一方面所述的信道反馈模型确定方法的步骤；或者所述计算机程序被处理器执行时实现上述第二方面所述的信道反馈模型确定方法的步骤。
本公开实施例中，由终端接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。这样，通过终端对目标信道反馈模型进行模型训练或选择目标信道反馈模型，能够降低网络侧的开销。
附图说明
为了更清楚地说明本公开实施例的技术方案，下面将对本公开实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本公开的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳 动性的前提下，还可以根据这些附图获得其他的附图。
图1是本公开实施例提供的一种信道反馈模型确定方法的流程图之一；
图2是本公开实施例提供的一种信道反馈模型的结构示意图之一；
图3是本公开实施例提供的一种信道反馈模型确定方法的流程图之二；
图4是本公开实施例提供的一种模型信息的结构示意图；
图5是本公开实施例提供的一种模型训练结果的示意图；
图6是本公开实施例提供的一种信道反馈模型的分类示意图；
图7a是本公开实施例提供的一种信道反馈模型的示意图之一；
图7b是本公开实施例提供的一种信道反馈模型的示意图之二；
图7c是本公开实施例提供的一种信道反馈模型的示意图之三；
图7d是本公开实施例提供的一种信道反馈模型的示意图之四；
图7e是本公开实施例提供的一种信道反馈模型的示意图之五；
图8是本公开实施例提供的一种信道反馈模型确定方法的流程图之三；
图9是本公开实施例提供的一种信道反馈模型确定方法的流程图之四；
图10是本公开实施例提供的一种信道反馈模型确定方法的流程图之五；
图11是本公开实施例提供的一种终端的结构示意图之一；
图12是本公开实施例提供的一种网络侧设备的结构示意图之一；
图13是本公开实施例提供的一种终端的结构示意图之二；
图14是本公开实施例提供的一种网络侧设备的结构示意图之二。
具体实施方式
下面将结合本公开实施例中的附图，对本公开实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本公开一部分实施例，而不是全部的实施例。基于本公开中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本公开保护的范围。
本公开实施例中，提出了一种信道反馈模型确定方法、终端及网络侧设备，以解决网络侧需要采集大量的信道数据对信道反馈模型进行训练，导致网络侧的开销较大的问题。
参见图1，图1是本公开实施例提供的一种信道反馈模型确定方法的流 程图，用于终端，如图1所示，所述方法包括以下步骤：
步骤101、接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息。
其中，信道状态信息参考信号(Channel state information Reference Signal，CSI-RS)由网络侧设备发送，用于终端进行信道测量，获取信道信息。信道信息可以为信道状态信息(Channel State Information，CSI)。
步骤102、基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型。
其中，目标信道反馈模型可以包括编码模型和解码模型。
需要说明的是，如图2所示，在对目标信道反馈模型进行模型训练的过程中，可以对解码模型的部分层进行冻结，仅对冻结的部分层外的其余层进行训练，冻结的层可以称为冻结层，进行训练的层可以称为微调层。一种实施方式中，在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变，所述目标层神经网络为所述解码模型的至少一层神经网络。
可以通过如下两种方式对目标信道反馈模型进行模型训练：
(1)由终端确定在线训练的训练配置。终端可以自身确定最佳的训练配置和需要向网络侧设备反馈的模型层数和权值系数。终端可以从解码模型的最后一层开始进行训练，直到性能没有明显提升。示例地，终端可以按优先级1训练倒数第一层，按照优先级2训练倒数第一层及倒数第二层，按照优先级3训练倒数第一层、倒数第二层及倒数第三层，…，按照优先级m训练倒数第一层、倒数第二层、倒数第三层，…，倒数第m层，直至训练精度不增加即停止训练，其中，优先级1为最高优先级。
(2)由网络侧设备确定在线训练的训练配置。网络侧设备可以根据历史大数据分析结果为终端配置较优的训练配置。可以将模型结构分为冻结层和微调层，对于冻结层，终端在模型训练进行反向传播时不进行权值系数的更新，对于微调层，终端在模型训练进行反向传播时通过梯度下降进行权值系 数的更新。终端只需反馈微调层的权值系数，可以降低训练开销和反馈开销。一种实施方式中，训练配置如下所示：微调层：1，2….m；冻结层：m…n；训练迭代次数。
步骤103、向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
其中，终端可以向所述网络侧设备发送进行模型训练后的目标信道反馈模型中的解码模型的模型信息。
一种实施方式中，如图3所示，当终端利用自身获取的信道信息对目标信道反馈模型进行在线训练后，可以在通过上行业务资源传输请求获取上行传输资源后，在服务数据适应协议(Service Data Adaptation Protocol，SDAP)数据包中传输目标信道反馈模型中的解码模型的模型信息，其中，如图4所示，该模型信息可以包括:模型标识(Identifier，ID)；需要更新的模型层数信息(例如，L1到Ln)；需要更新的模型权值系数。其中，上行业务资源传输请求可以包括：终端通过物理上行控制信道(Physical Uplink Control Channel，PUCCH)向无线接入网(Radio Access Network，RAN)发送调度请求(Scheduling Request，SR)信令，RAN通过物理下行控制信道(Physical downlink control channel，PDCCH)或下行控制信息(Downlink Control Information，DCI)向终端发送上行(Uplink，UL)允许(grant)信令，终端通过媒体接入控制(Medium Access Control，MAC)控制单元(Control Element，CE)向RAN发送缓存状态报告(Buffer Status Report，BSR)信令，RAN通过DCI或PDCCH向终端发送UL grant信令。
需要说明的是，网络侧设备(例如基站)可以周期性地发送CSI-RS；终端接收到CSI-RS后，进行信道估计，并使用目标信道反馈模型中的编码模型进行压缩量化后，向网络侧设备发送压缩量化后的信道信息；网络侧设备接收到压缩量化后的信道信息，采用更新的解码模型对信道信息进行解压，获取完整的下行信道估计信息。
需要说明的是，编码模型与解码模型在模型训练时可以采用端到端的方式统一训练，在部署时也配对使用；如果使用不匹配的编码模型与解码模型会造成明显的信道信息恢复精度下降。相关技术中模型训练由网络侧设备执 行，由于模型训练需要大量的算力，需要网络侧设备采集大量的信道数据，开销较大。并且，受限于训练数据的采集场景与采集规模的限制，由于终端所处的信道环境多变，固定不变的信道反馈模型很难在所有无线场景中都取得较好的性能，即存在泛化性问题。本公开实施例中，在终端对目标信道反馈模型进行模型训练，或基于信道信息选择目标信道反馈模型，能够解决由于编码模型与解码模型在不同环境下的泛化性带来的恢复精度下降的问题。
本公开实施例中，由终端接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。这样，通过终端对目标信道反馈模型进行模型训练或选择目标信道反馈模型，能够降低网络侧的开销。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
该实施方式中，通过终端向网络侧设备发送解码模型的模型标识，和/或，解码模型的全部权值系数或解码模型的部分权值系数，从而网络侧设备能够基于终端发送的模型标识和/或权值系数对解码模型进行更新。
可选地，所述基于所述信道信息对目标信道反馈模型进行模型训练，包括：
在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
示例地，目标信道反馈模型可以为所述网络侧设备采用训练样本对基础模型进行训练获得。如图2所示，基础模型中的编码模型由四层全连接层构成，基础模型中的解码模型也由四层全连接层构成。基础模型的初始权值系数可以是网络侧设备通过收集到的大数据进行训练得到，终端可以选取三种 不同大小的信道数据集对基础模型进行再训练。例如，如图5所示，可以通过大数据集、中数据集及小数据集对基础模型进行再训练。并且，可以冻结基础模型的不同层对基础模型进行再训练。从图5可以看出，在终端侧的再训练存在较优的训练策略，训练最后一层的归一化均方误差(Normalized Mean Square Error，NMSE)较高，但训练最后两层后NMSE快速下降，之后的边际效应递减，再训练最后两层是平衡效率和性能较优的选择。可以通过优化解码模型的部分层的权值系数来对解码模型的性能进行优化。对于终端而言，只反馈部分的解码模型的层数权值系数可以大大降低空口传输的开销。
该实施方式中，在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变，从而在模型训练时仅训练部分层的权值系数，进而向网络侧设备仅传递解码模型的部分层的权值系数，能够降低空口传输的开销。
可选地，所述方法还包括：
向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
所述基于所述信道信息对目标信道反馈模型进行模型训练，包括：
基于所述训练配置对目标信道反馈模型进行模型训练。
一种实施方式中，终端可以向网络侧设备发送上行探测参考信号(Sounding Reference Signal，SRS)，并接收所述网络侧设备发送的用于模型训练的训练配置，该训练配置可以是基于信道特征确定，该信道特征可以是基于终端发送的信道信息及所述SRS确定。
应理解，在无线通信系统中，无线信道的多径、衰落、时变及频率选择性等特性会对信息传输的性能带来巨大的影响。无线信道特征包括上下行的多径个数、角度及多径时延等参数，对于时分复用(Time Division Duplex，TDD)系统而言，无线信道具有较好的互异性，终端可以向网络侧设备发送用于信道信息获取的上行探测参考信号SRS，网络侧设备通过对SRS的测量， 获取上行信道矩阵H，并利用上下行互易性，得到下行信道矩阵H’，并通过下行信道矩阵H’计算用于下行数据传输的预编码矩阵及波束赋形矢量。对于频分复用(Frequency Division Duplex，FDD)系统，上下行部署在不同的频率上，并不存在完全的信道互易性，无线信道上各条径的角度和时延主要受收发端的空间角度关系、无线环境中反射体的位置和材质等因素决定，和无线信号的频率并不存在很强的关系。而无线信道的频率会很大程度上影响无线信道的路径损耗、穿透损耗及极化泄露因子等成分，因而会对各条径经历的幅度变化和相位变化产生较大的影响。因此，FDD系统中上下行信道的频域基矢量和空域基矢量存在较强的相关性，而加权系数，代表无线信道中各条径经历的幅度变化和相位变化，较为独立，即存在部分互易性。
一种实施方式中，网络侧设备可以根据信道的部分互易性获取下行信道的部分信道特征，同时也可以根据终端反馈的信道信息来获取信道特征。
该实施方式中，终端向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；终端基于所述训练配置对目标信道反馈模型进行模型训练，从而能够通过网络侧设备确定训练配置。
可选地，所述基于所述信道信息对目标信道反馈模型进行模型训练之前，所述方法还包括：
基于所述信道信息选择目标信道反馈模型。
可选地，所述基于所述信道信息选择目标信道反馈模型，包括：
基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或动态环境参数；
基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
其中，静态特征参数可以包括但不限于：网络配置参数，例如宏站/小站、网络侧多天线端口数等；终端能力参数，例如终端多天线端口数、终端人工智能(Artificial Intelligence，AI)推理能力等。上述静态特征参数一般属于终端初始接入网络时需要上报的信息，且在通信业务发生过程中基本保持不变。动态环境参数可以对应各类动态场景，动态环境参数可以包括信道环境参数及信道质量参数等，信道环境参数可以为视距或非视距传输下的信道环境参 数，信道质量参数可以包括信噪比范围等。
一种实施方式中，网络侧设备可以在对基础模型进行训练的阶段，基于目标参数形成不同训练集，并在不同训练集的基础上训练多个模型，将多个模型按不同适用场景分为多个类别。
一种实施方式中，终端可以从多个模型中选择目标信道反馈模型，如图6所示，该多个模型可以先按静态特征参数分类，得到多个半静态类别，再由静态特征参数决定的每一子类中继续按照动态环境参数进行分类，每一子类可以划分多个动态子类模型。
需要说明的是，每个子类内的多个模型之间可以存在编码模型或解码模型共享。编码模型可以称为编码器，解码模型可以称为解码器。如图7a至图7e所示，列举了5种编码模型-解码模型的模式，可以多个编码模型共享一个解码模型，或者多个解码模型共享一个编码模型。对于所列的5种模式，可按照是否存在多个可供选择的编码模型或解码模型分为两类，即acde/b，或者abde/c。每一对编码模型-解码模型对应一组动态类别，在实际使用时，可以选择某一组具体的编码模型-解码模型来进行CSI的压缩与恢复。
一种实施方式中，基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型，可以是，对于每对编码模型-解码模型标注适用的静态特征参数和/或动态环境参数，根据目标参数按查找字典的方式选择对应的编码模型-解码模型。
一种实施方式中，至少两个信道反馈模型可以按照适用的静态特征参数和/或动态环境参数划分为多个子类，基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型，可以是，根据目标参数按查找字典的方式选择对应子类的编码模型-解码模型，依次使用选择的子类内的编码模型-解码模型对信道信息进行压缩恢复，并计算恢复精度，遍历子类内的模型后按恢复精度排序，选择恢复精度最高的模型作为目标信道反馈模型。
另外，在选择目标信道反馈模型后，终端可以将选择的目标信道反馈模型的解码模型通知网络侧设备，通知方式可以为发送携带解码模型的信令至网络侧设备，当编码模型或解码模型存在模型共享甚至某一子类别中仅存在单个编码模型或解码模型时，可以不用同步解码模型，是否通知对端选择的 目标信道反馈模型的解码模型由网络侧设备指定。
一种实施方式中，可以通过对无线资源控制(Radio Resource Control，RRC)或者MAC CE信令扩展实现解码模型的模型信息的传递，以MAC CE信令实现解码模型的模型信息的传递为例，可以在MAC CE信令中增加逻辑信道组(Logic Channel Group，LCG)ID字段，通过该字段携带解码模型的模型标识。
该实施方式中，基于静态特征参数和/或动态环境参数从至少两个信道反馈模型中选择目标信道反馈模型，能够获取与当前静态特征参数和/或动态环境参数较为匹配的信道反馈模型。
参见图8，图8是本公开实施例提供的一种信道反馈模型确定方法的流程图，用于网络侧设备，如图8所示，所述方法包括以下步骤：
步骤201、向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息。
步骤202、接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
需要说明的是，本实施例作为与图1所示的实施例中对应的网络侧设备的实施方式，其具体的实施方式可以参见图1所示的实施例中的相关说明，为避免重复说明，本实施例不再赘述。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述接收所述终端发送的目标信道反馈模型中的解码模型的模型信息之前，所述方法还包括:
接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
上述可选的实施方式可以参见图1所示的实施例中的相关说明，为避免重复说明，本实施例不再赘述。
可选地，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
所述基于所述信道信息向所述终端发送用于模型训练的训练配置之前，所述方法还包括：
基于所述信道信息获取第一信道特征；
将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
该实施方式中，基于所述信道信息获取第一信道特征；将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。这样，能够通过训练基础模型的训练样本确定的平均信道特征与第一信道特征的比对确定训练配置，能够对不同的第一信道特征确定适宜的训练配置，从而能够获得较好的模型训练效果。
可选地，所述第一信道特征包括如下至少一项：
角度域信息；
时延域信息；
多普勒域信息。
其中，角度域信息反映的是波束/天线的稀疏性，例如，可以通过角度域信息指示的前3个路径增益最大的波束在整体功率增益中的占比是否超过第一门限确定不同的训练配置，第一门限通过第二信道特征确定，一种实施方式中，第二信道特征中角度域信息指示的前3个路径增益最大的波束在整体功率增益中的占比为第一门限，示例地，第一门限可以为80％。
另外，时延域信息反映的是时延域/多径的稀疏性，非视距(Non Line of Sight，NLOS)或者视距(Line of Sight，LOS)信道，例如，可以通过时延域信息指示的K因子(即主径占据的功率比)是否超过第二门限确定不同的训练配置，第二门限通过第二信道特征确定，一种实施方式中，第二信道特征中时延域信息指示的K因子为第二门限，示例地，第二门限可以为50％。
另外，多普勒域信息反映的是反映的是时间域上的稀疏性，例如，可以 通过多普勒域信息指示的前5个主径的频偏功率占比是否超过第三门限确定不同的训练配置，第三门限通过第二信道特征确定，一种实施方式中，第三信道特征中多普勒域信息指示的前5个主径的频偏功率占比为第三门限，示例地，第三门限可以为60％。
一种实施方式中，通过角度域信息指示的前3个路径增益最大的波束在整体功率增益中的占比为A，通过时延域信息指示的K因子为B，通过多普勒域信息指示的前5个主径的频偏功率占比为C。如图2所示，编码模型由四层全连接层构成，解码模型由四层全连接层构成。模型结构分为冻结层和微调层，对于冻结层，终端在模型训练进行反向传播时不进行权值系数的更新，对于微调层，终端在模型训练进行反向传播时通过梯度下降进行权值系数的更新。
在A>80％或者B>50％时，用于模型训练的训练配置可以为：冻结层的层数设置的较多，微调层为解码模型的第四层；
在A>50％或者B>50％时，用于模型训练的训练配置可以为：冻结层的层数设置的较少，微调层为解码模型的第三层及第四层；
在C>60％时，用于模型训练的训练配置可以为：连续样本数量少，样本数量100，持续时间500ms；
在C<60％时，用于模型训练的训练配置可以为：连续样本数量多，样本数量1000，持续时间1s。
以下通过两个具体的实施例对本公开实施例的信道反馈模型确定方法进行说明：
实施例1：
以网络侧设备为基站为例，通过终端对编码模型和解码模型进行在线训练实现信道反馈模型的确定，具体的，如图9所示，包括如下过程：
(1)部署信道反馈模型，该信道反馈模型包括编码模型和解码模型，示例地，可以是基站向终端发送信道反馈模型，或者可以在终端预先配置信道反馈模型；
(2)基站通过RRC重配消息发送CSI-RS配置；
(3)基站按照CSI-RS配置向终端周期性发送CSI-RS；
(4)终端接收CSI-RS，并测量CSI-RS获取信道状态信息；
终端向基站发送CSI报告，并向基站发送SRS；
(5)基站根据信道特征选择在线训练的配置；
(6)基站将在线训练的配置发送给终端；
(7)终端基于信道状态信息对编码模型和解码模型进行在线训练；
(8)终端将训练的解码模型的模型信息发送给基站；
(9)基站基于接收到的模型信息更新解码模型；
(10)基站向终端周期性发送CSI-RS；
(11)终端使用训练的编码模型压缩CSI，并对压缩的CSI进行量化处理；
(12)终端向基站上报压缩量化后的CSI；
(13)基站使用更新后的解码模型对CSI进行解码。
实施例2：
以网络侧设备为基站为例，通过终端选择信道反馈模型实现信道反馈模型的确定，具体的，如图10所示，包括如下过程：
(1)部署信道反馈模型，该信道反馈模型包括编码模型和解码模型，示例地，可以是基站向终端发送信道反馈模型，或者可以在终端预先配置信道反馈模型；
(2)基站通过RRC重配消息发送CSI-RS配置；
(3)基站按照CSI-RS配置向终端周期性发送CSI-RS；
(4)终端接收CSI-RS，并测量CSI-RS获取信道状态信息；
(5)终端根据测量获取的信道状态信息选择解码模型；
(6)终端将选择的解码模型的模型信息发给基站；
(7)基站基于接收到的模型信息更新解码模型；
(8)基站向终端周期性发送CSI-RS；
(9)终端使用编码模型压缩CSI，并对压缩的CSI进行量化处理；
(10)终端向基站上报压缩量化后的CSI；
(11)基站使用更新后的解码模型对CSI进行解码。
参见图11，图11是本公开实施例提供的一种终端的结构示意图，如图 11所示，终端300包括：
接收模块301，用于接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
处理模块302，用于基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
第一发送模块303，用于向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述处理模块具体用于：
在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
可选地，所述装置还包括：
第二发送模块，用于向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
所述处理模块具体用于：
基于所述训练配置对目标信道反馈模型进行模型训练。
可选地，所述装置还包括：
选择模块，用于基于所述信道信息选择目标信道反馈模型。
可选地，所述处理模块具体用于：
基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或动态环境参数；
基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
终端能够实现图1所示的方法实施例实现的各个过程，且能达到相同的有益效果，为避免重复，这里不再赘述。
参见图12，图12是本公开实施例提供的一种网络侧设备的结构示意图，如图12所示，网络侧设备400包括：
第一发送模块401，用于向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
接收模块402，用于接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述装置还包括:
第二发送模块，用于接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
可选地，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
所述装置还包括：
获取模块，用于基于所述信道信息获取第一信道特征；
确定模块，用于将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
可选地，所述第一信道特征包括如下至少一项：
角度域信息；
时延域信息；
多普勒域信息。
网络侧设备能够实现图8所示的方法实施例实现的各个过程，且能达到相同的有益效果，为避免重复，这里不再赘述。
本公开实施例还提供了一种终端，包括：处理器、存储器及存储在所述 存储器上并可在所述处理器上运行的程序，所述程序被所述处理器执行时实现上述信道反馈模型确定方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
具体的，参见图13所示，本公开实施例还提供了一种终端，包括总线501、收发机502、天线503、总线接口504、处理器505和存储器506。
其中，所述收发机502，用于接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
所述处理器505，用于基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
所述收发机502，还用于向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述处理器505用于：
在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
可选地，所述收发机502还用于：向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
所述处理器505还用于：
基于所述训练配置对目标信道反馈模型进行模型训练。
可选地，所述处理器505还用于：基于所述信道信息选择目标信道反馈模型。
可选地，所述处理器505还用于：
基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或动态环境参数；
基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
在图13中，总线架构(用总线501来代表)，总线501可以包括任意数量的互联的总线和桥，总线501将包括由处理器505代表的一个或多个处理器和存储器506代表的存储器的各种电路链接在一起。总线501还可以将诸如外围设备、稳压器和功率管理电路等之类的各种其他电路链接在一起，这些都是本领域所公知的，因此，本文不再对其进行进一步描述。总线接口504在总线501和收发机502之间提供接口。收发机502可以是一个元件，也可以是多个元件，比如多个接收器和发送器，提供用于在传输介质上与各种其他装置通信的单元。经处理器505处理的数据通过天线503在无线介质上进行传输，进一步，天线503还接收数据并将数据传送给处理器505。
处理器505负责管理总线501和通常的处理，还可以提供各种功能，包括定时，外围接口，电压调节、电源管理以及其他控制功能。而存储器506可以被用于存储处理器505在执行操作时所使用的数据。
可选的，处理器505可以是中央处理器(Central Processing Unit，CPU)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程逻辑门阵列(Field Programmable Gate Array，FPGA)或复杂可编程逻辑器件(Complex Programmable logic device，CPLD)。
本公开实施例还提供了一种网络侧设备，包括：处理器、存储器及存储在所述存储器上并可在所述处理器上运行的程序，所述程序被所述处理器执行时实现上述信道反馈模型确定方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
具体的，参见图14所示，本公开实施例还提供了一种网络侧设备，包括总线601、收发机602、天线603、总线接口604、处理器605和存储器606。
其中，所述收发机602，用于向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
所述收发机602，还用于接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
可选地，所述模型信息包括如下至少一项：
所述解码模型的模型标识；
所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
可选地，所述收发机602，还用于：
接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
可选地，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
所述处理器605用于：
基于所述信道信息获取第一信道特征；
将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
可选地，所述第一信道特征包括如下至少一项：
角度域信息；
时延域信息；
多普勒域信息。
在图14中，总线架构(用总线601来代表)，总线601可以包括任意数量的互联的总线和桥，总线601将包括由处理器605代表的一个或多个处理器605和存储器606代表的存储器的各种电路链接在一起。总线601还可以将诸如外围设备、稳压器和功率管理电路等之类的各种其他电路链接在一起，这些都是本领域所公知的，因此，本文不再对其进行进一步描述。总线接口604在总线601和收发机602之间提供接口。收发机602可以是一个元件，也可以是多个元件，比如多个接收器和发送器，提供用于在传输介质上与各种其他装置通信的单元。经处理器605处理的数据通过天线603在无线介质上进行传输，进一步，天线603还接收数据并将数据传送给处理器605。
处理器605负责管理总线601和通常的处理，还可以提供各种功能，包括定时，外围接口，电压调节、电源管理以及其他控制功能。而存储器606 可以被用于存储处理器605在执行操作时所使用的数据。
可选的，处理器605可以是CPU、ASIC、FPGA或CPLD。
本公开实施例还提供一种计算机可读存储介质，计算机可读存储介质上存储有计算机程序，该计算机程序被处理器执行时实现上述信道反馈模型确定方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。其中，所述的计算机可读存储介质，如只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本公开的技术方案本质上或者说对相关技术做出贡献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(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本公开各个实施例所述的方法。
上面结合附图对本公开的实施例进行了描述，但是本公开并不局限于上述的具体实施方式，上述的具体实施方式仅仅是示意性的，而不是限制性的，本领域的普通技术人员在本公开的启示下，在不脱离本公开宗旨和权利要求所保护的范围情况下，还可做出很多形式，均属于本公开的保护之内。

Claims (17)

  1. 一种信道反馈模型确定方法，应用于终端，所述方法包括：
    接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
    基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
    向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
  2. 根据权利要求1所述的方法，其中，所述模型信息包括如下至少一项：
    所述解码模型的模型标识；
    所述解码模型的全部权值系数，或所述解码模型的部分权值系数。
  3. 根据权利要求1所述的方法，其中，所述基于所述信道信息对目标信道反馈模型进行模型训练，包括：
    在对目标信道反馈模型进行模型训练的过程中，基于所述信道信息对所述解码模型中的目标层神经网络的权值系数进行训练，并保持所述目标信道反馈模型中编码模型的权值系数及所述解码模型中除所述目标层神经网络外的其他层神经网络的权值系数不变；
    其中，所述目标层神经网络为所述解码模型的至少一层神经网络。
  4. 根据权利要求1所述的方法，所述方法还包括：
    向所述网络侧设备发送所述信道信息，并接收所述网络侧设备基于所述信道信息发送的用于模型训练的训练配置；
    所述基于所述信道信息对目标信道反馈模型进行模型训练，包括：
    基于所述训练配置对目标信道反馈模型进行模型训练。
  5. 根据权利要求1所述的方法，其中，所述基于所述信道信息对目标信道反馈模型进行模型训练之前，所述方法还包括：
    基于所述信道信息选择目标信道反馈模型。
  6. 根据权利要求1或5所述的方法，其中，所述基于所述信道信息选择目标信道反馈模型，包括：
    基于所述信道信息确定目标参数，所述目标参数包括静态特征参数和/或 动态环境参数；
    基于所述目标参数从至少两个信道反馈模型中选择目标信道反馈模型。
  7. 一种信道反馈模型确定方法，应用于网络侧设备，所述方法包括：
    向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
    接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
  8. 根据权利要求7所述的方法，其中，所述接收所述终端发送的目标信道反馈模型中的解码模型的模型信息之前，所述方法还包括:
    接收所述终端发送的所述信道信息，并基于所述信道信息向所述终端发送用于模型训练的训练配置；
    其中，所述训练配置用于对所述目标信道反馈模型进行模型训练。
  9. 根据权利要求8所述的方法，其中，所述目标信道反馈模型为所述网络侧设备采用训练样本对基础模型进行训练获得；
    所述基于所述信道信息向所述终端发送用于模型训练的训练配置之前，所述方法还包括：
    基于所述信道信息获取第一信道特征；
    将所述第一信道特征与第二信道特征进行比对，并基于比对结果确定用于模型训练的训练配置，所述第二信道特征为基于所述训练样本确定的平均信道特征。
  10. 根据权利要求9所述的方法，其中，所述第一信道特征包括如下至少一项：
    角度域信息；
    时延域信息；
    多普勒域信息。
  11. 一种终端，包括：
    接收模块，用于接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
    处理模块，用于基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
    第一发送模块，用于向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
  12. 一种网络侧设备，包括：
    第一发送模块，用于向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
    接收模块，用于接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
  13. 一种终端，包括收发机和处理器，
    所述收发机，用于接收网络侧设备发送的信道状态信息参考信号，对所述信道状态信息参考信号进行信道测量获取信道信息；
    所述处理器，用于基于所述信道信息对目标信道反馈模型进行模型训练，或基于所述信道信息选择目标信道反馈模型；
    所述收发机，还用于向所述网络侧设备发送所述目标信道反馈模型中的解码模型的模型信息。
  14. 一种网络侧设备，包括收发机和处理器，
    所述收发机，用于向终端发送信道状态信息参考信号，以使所述终端对所述信道状态信息参考信号进行信道测量获取信道信息；
    所述收发机，还用于接收所述终端发送的目标信道反馈模型中的解码模型的模型信息，其中，所述目标信道反馈模型为所述终端基于所述信道信息进行模型训练获得，或所述目标信道反馈模型为所述终端基于所述信道信息选择。
  15. 一种终端，包括：处理器、存储器及存储在所述存储器上并可在所述处理器上运行的程序，其中，所述程序被所述处理器执行时实现如权利要求1至6中任一项所述的信道反馈模型确定方法的步骤。
  16. 一种网络侧设备，包括：处理器、存储器及存储在所述存储器上并可在所述处理器上运行的程序，其中，所述程序被所述处理器执行时实现如 权利要求7至10中任一项所述的信道反馈模型确定方法的步骤。
  17. 一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机程序，其中，所述计算机程序被处理器执行时实现如权利要求1至6中任一项所述的信道反馈模型确定方法的步骤；或者所述计算机程序被处理器执行时实现如权利要求7至10中任一项所述的信道反馈模型确定方法的步骤。
PCT/CN2023/102375 2022-06-28 2023-06-26 信道反馈模型确定方法、终端及网络侧设备 Ceased WO2024002003A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210748510.1 2022-06-28
CN202210748510.1A CN117353849A (zh) 2022-06-28 2022-06-28 信道反馈模型确定方法、终端及网络侧设备

Publications (1)

Publication Number Publication Date
WO2024002003A1 true WO2024002003A1 (zh) 2024-01-04

Family

ID=89356246

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/102375 Ceased WO2024002003A1 (zh) 2022-06-28 2023-06-26 信道反馈模型确定方法、终端及网络侧设备

Country Status (2)

Country Link
CN (1) CN117353849A (zh)
WO (1) WO2024002003A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
USD1115233S1 (en) * 2024-06-19 2026-03-03 Sanliuyidu (China) Co., Ltd Shoe

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118368210B (zh) * 2024-05-16 2025-06-24 鹏城实验室 模型动态更新方法、设备、存储介质及计算机程序产品

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021108940A1 (en) * 2019-12-01 2021-06-10 Nokia Shanghai Bell Co., Ltd. Channel state information feedback
US20210273707A1 (en) * 2020-02-28 2021-09-02 Qualcomm Incorporated Neural network based channel state information feedback
CN113810086A (zh) * 2020-06-12 2021-12-17 华为技术有限公司 信道信息反馈方法、通信装置及存储介质
US20220060917A1 (en) * 2020-08-18 2022-02-24 Qualcomm Incorporated Online training and augmentation of neural networks for channel state feedback
WO2022040046A1 (en) * 2020-08-18 2022-02-24 Qualcomm Incorporated Reporting configurations for neural network-based processing at a ue

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021108940A1 (en) * 2019-12-01 2021-06-10 Nokia Shanghai Bell Co., Ltd. Channel state information feedback
US20210273707A1 (en) * 2020-02-28 2021-09-02 Qualcomm Incorporated Neural network based channel state information feedback
CN113810086A (zh) * 2020-06-12 2021-12-17 华为技术有限公司 信道信息反馈方法、通信装置及存储介质
US20220060917A1 (en) * 2020-08-18 2022-02-24 Qualcomm Incorporated Online training and augmentation of neural networks for channel state feedback
WO2022040046A1 (en) * 2020-08-18 2022-02-24 Qualcomm Incorporated Reporting configurations for neural network-based processing at a ue

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
USD1115233S1 (en) * 2024-06-19 2026-03-03 Sanliuyidu (China) Co., Ltd Shoe

Also Published As

Publication number Publication date
CN117353849A (zh) 2024-01-05

Similar Documents

Publication Publication Date Title
US11522597B2 (en) Beam information feedback method and apparatus, and configuration information feedback method and apparatus
CN111756457B (zh) 信道预测方法、装置及计算机存储介质
US12119909B2 (en) Adaptive CSI reporting and PRB bundling in AAS
CN117955595B (zh) 信道状态信息报告配置的传输方法、装置及存储介质
CN116996189A (zh) 通信方法及装置、芯片、芯片模组、存储介质
WO2024002003A1 (zh) 信道反馈模型确定方法、终端及网络侧设备
WO2023195891A1 (en) Methods for dynamic channel state information feedback reconfiguration
WO2025035776A1 (zh) 信道状态信息的发送和接收方法、通信装置、及存储介质
WO2025185162A1 (zh) 信道状态信息报告的传输方法、通信装置以及存储介质
WO2022009151A1 (en) Shared csi-rs for partial-reciprocity based csi feedback
CN121173444B (zh) 通信方法以及相关装置
US20240056833A1 (en) Predictive beam management with per-beam error statistics
CN118449878A (zh) 一种信息处理方法、设备及可读存储介质
WO2025209453A1 (zh) 通信方法、装置、终端、网络侧设备、介质及产品
US20250096871A1 (en) Event-Triggered CSI Reporting
US20250357984A1 (en) Information sending method, information receiving method, communication device, and storage medium
WO2025107651A1 (zh) 信道状态信息报告的发送方法、接收方法及装置
WO2024207736A1 (zh) 数据集发送和接收方法、装置及存储介质
JP2026513338A (ja) 通信方法及び通信装置
WO2025035777A1 (zh) 性能指示的发送和接收方法、通信装置、及存储介质
WO2024065810A1 (en) Method for uplink sounding reference signal precoder selection for interference suppression
US12597984B2 (en) Methods, apparatus and machine-readable media relating to channel estimation
WO2024164638A1 (zh) 指示信息发送方法、指示信息接收方法、装置及存储介质
EP4632618A1 (en) Data set sending method and receiving method, and communication device and storage medium
US20260107188A1 (en) Information compression method and apparatus, and terminal device and network device

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: 23830180

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 10/04/2025)

122 Ep: pct application non-entry in european phase

Ref document number: 23830180

Country of ref document: EP

Kind code of ref document: A1