WO2022252162A1 - 一种模型训练方法、模型训练装置及存储介质 - Google Patents

一种模型训练方法、模型训练装置及存储介质 Download PDF

Info

Publication number
WO2022252162A1
WO2022252162A1 PCT/CN2021/098008 CN2021098008W WO2022252162A1 WO 2022252162 A1 WO2022252162 A1 WO 2022252162A1 CN 2021098008 W CN2021098008 W CN 2021098008W WO 2022252162 A1 WO2022252162 A1 WO 2022252162A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
access network
training
layer
data
Prior art date
Application number
PCT/CN2021/098008
Other languages
English (en)
French (fr)
Inventor
牟勤
洪伟
赵中原
王靖壹
Original Assignee
北京小米移动软件有限公司
北京邮电大学
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 北京小米移动软件有限公司, 北京邮电大学 filed Critical 北京小米移动软件有限公司
Priority to PCT/CN2021/098008 priority Critical patent/WO2022252162A1/zh
Priority to CN202180001782.9A priority patent/CN115735214A/zh
Publication of WO2022252162A1 publication Critical patent/WO2022252162A1/zh

Links

Images

Classifications

    • 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

Definitions

  • the present disclosure relates to the technical field of wireless communication, and in particular, to a model training method, a model training device, and a storage medium.
  • the wireless network AI architecture provides the basis for the realization of wireless artificial intelligence.
  • the continuity of the AI analysis service obtained by the terminal and the wireless Artificial intelligence performs mobility management, and regulates and optimizes the AI architecture of the wireless network.
  • 3rd Generation Partnership Project 3rd Generation Partnership Project, 3GPP
  • a wireless network architecture supporting artificial intelligence was proposed to obtain an artificial intelligence wireless network enabled by big data.
  • the wireless network architecture supporting artificial intelligence can handle multiple training task scenarios at the same time.
  • the model needs to be trained separately for each training task, resulting in relatively large training overhead and reducing data security risks.
  • the present disclosure provides a model training method, a model training device and a storage medium.
  • a model training method which is applied to an operation and maintenance management OAM entity, and the method includes:
  • the radio access network device group includes a first number of radio access network devices; determining the first number The first quantity model training structure corresponding to the radio access network device, and based on the first quantity model training structure, determine the first quantity unique model layer; send the first quantity model layer to the first quantity radio access network device Quantity-specific structural parameters of the model layers.
  • the determining the first number of model training structures corresponding to the first number of radio access network devices includes:
  • the determining the first number of unique model layers based on the first number of model training structures includes:
  • the method further includes:
  • the data identification corresponding to the access network device classifying and processing all the data with the data identification to obtain model training data and model label values; using the model training data as the first input data and inputting it into the shared model
  • the layer obtains the first output data output by the shared model layer; and sends the model tag value and the first output data to the multiple radio access network devices.
  • the method further includes:
  • the updating of the shared model layer based on the training loss value includes:
  • Weighting the training loss value to obtain a weighted loss value determining the current model parameters and model learning rate of the shared model layer; based on the weighted loss value, model parameters and model learning rate, determining the shared model layer updating parameters, and updating the structural parameters of the shared model layer based on the updating parameters.
  • the method after updating the shared model layer based on the update parameters, the method includes:
  • the T is the preset number of times to update the shared model layer and the unique model layer, and the shared model layer structure parameters are used by the wireless access network device to synthesize wireless access A model for network device subscriptions.
  • the grouping the radio access network devices that send the model subscription request to obtain at least one radio access network device group includes:
  • the method further includes:
  • a model training method which is applied to a radio access network device, and the method includes:
  • the unique model layer is determined by OAM by dividing the first quantity model training structure; the first quantity model training structure is the first quantity model training structure included by OAM based on the radio access network equipment group A number of model subscription requests for radio access network devices is determined.
  • the method further includes:
  • the determining a training loss value based on the model training data and the second output data includes:
  • the method further includes:
  • a model training device which is characterized in that it is applied to an OAM entity, and the device includes:
  • a grouping module configured to group multiple wireless access network devices that send model subscription requests to obtain at least one wireless access network device group, where the wireless access network device group includes a first number of wireless access network devices; determine A module for determining a first quantity model training structure corresponding to the first quantity of radio access network devices, and based on the first quantity model training structure, determining a first quantity of unique model layers; a sending module for sending to all The first number of radio access network devices sends the structural parameters of the first number of unique model layers.
  • the determining module is configured to:
  • the determining module is configured to:
  • the determination module is also used for:
  • the data identification corresponding to the access network device classifying and processing all the data with the data identification to obtain model training data and model label values; using the model training data as the first input data and inputting it into the shared model
  • the layer obtains the first output data output by the shared model layer; and sends the model tag value and the first output data to the multiple radio access network devices.
  • the device further includes: an update module;
  • the update module is configured to update the structural parameters of the shared model layer based on the training loss values in response to receiving the training loss values sent by the plurality of wireless access network devices.
  • the update module is used for:
  • Weighting the training loss value to obtain a weighted loss value determining the current model parameters and model learning rate of the shared model layer; based on the weighted loss value, model parameters and model learning rate, determining the shared model layer updating parameters, and updating the structural parameters of the shared model layer based on the updating parameters.
  • the updating module is also used for:
  • the T is the preset number of times to update the shared model layer and the unique model layer, and the shared model layer structure parameters are used by the wireless access network device to synthesize wireless access A model for network device subscriptions.
  • the grouping module is used for:
  • the updating module is also used for:
  • a model training device which is characterized in that it is applied to radio access network equipment, and the device includes:
  • the receiving module is used to receive the structural parameters of the unique model layer sent by OAM; wherein, the unique model layer is determined by dividing the first quantity model training structure by OAM; the first quantity model training structure is OAM based on the wireless access network Determined by the model subscription request of the first quantity of radio access network devices included in the device group.
  • the receiving module is also used for:
  • the device further includes: a determination module
  • a determination module configured to determine, among the model label values, a model label value corresponding to the radio access network device based on the identifier carried in the model training data; and calculate the second output data and the training label value, determine a training loss value, and update the structural parameters of the unique model layer based on the training loss value.
  • the receiving module is also used for:
  • a model training device including:
  • a processor a memory for storing processor-executable instructions; wherein the processor is configured to: execute the first aspect or the model training method described in any one of the implementations of the first aspect, or execute the second aspect Or the model training method described in any one of the implementation manners in the second aspect.
  • a non-transitory computer-readable storage medium When the instructions in the storage medium are executed by the processor of the mobile terminal, the mobile terminal can execute the first aspect or the first The model training method described in any one of the implementation manners of the second aspect, or enabling the mobile terminal to execute the second aspect or the model training method described in any one of the implementation manners of the second aspect.
  • OAM needs to carry out data identification on the model training data.
  • the identification information includes the gNB-CU to which the training data belongs and the data ID, etc.
  • the gNB-CU information is identified so that each gNB-CU can filter out the information belonging to the gNB.
  • -CU training data so as to use this part of the data to update its unique model layer, and mark the data ID information so that each gNB-CU can retrieve the label value and other information of the data according to the ID information, and use it to calculate the training loss value .
  • Fig. 1 is a schematic diagram of a wireless network architecture supporting artificial intelligence according to an exemplary embodiment.
  • Fig. 2 is a schematic diagram showing a multi-task model training task according to an exemplary embodiment.
  • Fig. 3 is a schematic diagram showing another multi-task model training task according to an exemplary embodiment.
  • Fig. 4 is a schematic diagram of a system architecture of a model training method according to an exemplary embodiment.
  • Fig. 5 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 6 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 7 is a schematic flowchart of a training model structure of a model training method according to an exemplary embodiment.
  • Fig. 8 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 9 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 10 is a schematic flowchart of collecting and identifying model training data for a training model structure according to an exemplary embodiment.
  • Fig. 11 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 12 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 13 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 14 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 15 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 16 is a flow chart of adding a radio access network device according to a model training method according to an exemplary embodiment.
  • Fig. 17 is a flowchart of a model training method according to an exemplary embodiment.
  • Fig. 18 is a flow chart of a radio access network device exiting a model training method according to an exemplary embodiment.
  • Fig. 19 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 20 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 21 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 22 is a flowchart showing a method for updating a model-specific layer parameter according to an exemplary embodiment.
  • Fig. 23 is a flow chart showing another model training method according to an exemplary embodiment.
  • Fig. 24 is a schematic diagram of model training and model inference of a model training method according to an exemplary embodiment.
  • Fig. 25 is a flow chart showing a model training method and a model reasoning method according to an exemplary embodiment.
  • Fig. 26 is a schematic diagram of a model training protocol and interface of a model training method according to an exemplary embodiment.
  • Fig. 27 is a schematic diagram of the protocol and interface of model inference in a model training method according to an exemplary embodiment.
  • Fig. 28 is a schematic diagram of a protocol and an interface for collecting model training data in a model training method according to an exemplary embodiment.
  • Fig. 29 is a schematic diagram of a protocol and interface for model inference data collection in a model training method according to an exemplary embodiment.
  • Fig. 30 is a block diagram of a model training device according to an exemplary embodiment.
  • Fig. 31 is a block diagram of another model training device according to an exemplary embodiment.
  • Fig. 32 is a block diagram showing a device for model training according to an exemplary embodiment.
  • Fig. 33 is a block diagram showing another apparatus for model training according to an exemplary embodiment.
  • the wireless network AI architecture provides the basis for the realization of wireless artificial intelligence.
  • the continuity of the AI analysis service obtained by the terminal, and the wireless Artificial intelligence performs mobility management, and regulates and optimizes the AI architecture of the wireless network.
  • a wireless network architecture supporting artificial intelligence was proposed to obtain an artificial intelligence wireless network enabled by big data.
  • Fig. 1 is a schematic diagram of a wireless network architecture supporting artificial intelligence according to an exemplary embodiment.
  • the wireless network architecture supporting artificial intelligence includes data collection/preparation unit, model training unit, model inference unit, action unit, training data, model performance feedback, model deployment/update, inference data, inference results and performance feedback.
  • Data collection/preparation unit collect data related to AI model training, updating, and reasoning, and preprocess the data according to the requirements of AI model training, updating, and reasoning for data content, size, format, and cycle
  • the data is provided to the model training unit and the model prediction unit as required.
  • the data collection/preparation unit will also judge the validity of the current AI model based on the collected data, and provide model performance feedback to the model training unit.
  • Model training unit responsible for training and updating the AI model.
  • the input data required for training and updating is provided by the data collection/preparation unit, and the trained or updated AI model is provided to the model reasoning unit.
  • Model reasoning unit Based on the AI model provided by the model training unit and the input data provided by the data collection/preparation unit, perform specific wireless network reasoning tasks, and provide the reasoning results to the action unit and data collection/preparation unit.
  • Action unit According to the reasoning results provided by the model reasoning unit, the corresponding network behavior is executed.
  • the action unit collects data on the network side and provides it to the data collection/preparation unit in the form of performance feedback.
  • Training data The data collection/preparation unit preprocesses the collected data, and provides the data required for AI model training and updating to the model training unit.
  • Model performance feedback The data collection/preparation unit judges the effectiveness of the current AI model based on the collected data (for example, comparing predicted data with actual measurement data), and provides model performance feedback to the model training unit.
  • Model deployment/update The model training unit provides the trained and updated AI model to the model reasoning unit.
  • the data collection/preparation unit preprocesses the collected data and provides the data required for AI model reasoning to the model reasoning unit.
  • the inference result generated according to the AI model is provided by the model inference unit to the action unit and data collection/preparation unit.
  • Performance feedback After performing the corresponding network behavior, the action unit will collect the data on the network side and provide it to the data collection/preparation unit.
  • FIG. 2 is a schematic diagram showing a multi-task model training task according to an exemplary embodiment. As shown in Figure 2, based on the two determined model subscription requests, first determine two training data including training data 1 and training data 2, train model 1 based on training data 1, train model 2 based on training data 2, and obtain the task 1 model Prediction and Model 2 Task Prediction. During this process, model information is not shared between different training tasks.
  • the terminal initiates a model subscription request to the wireless access network equipment of the 5G base station distributed unit (next Generation Node B Distributed Unit, gNB-DU), and the gNB-DU sends the model subscription request of the terminal to the 5G base station control unit (next Generation Node B Control Unit, gNB-CU), gNB-CU reports the model subscription request of the terminal to OAM; OAM selects an appropriate training model according to the model subscription request of the terminal, and collects model training data for model training; after the model training is completed, OAM Send the training model to the gNB-CU, and the gNB-CU collects model inference data and performs model inference; after the model inference is completed, the gNB-CU sends the inference model to the gNB-DU, and the gNB-DU sends the inference model to the terminal; the terminal according to The inference model executes the corresponding policy.
  • OAM In the process of model training, OAM needs gNB-CU to upload data to OAM, which poses a challenge to data security. If part of the model training work is transferred to gNB-CU, the uploading can be reduced. reduce data security risks.
  • model training work is all completed by OAM, which will consume more OAM computing resources. If part of the model training work is transferred to the gNB-CU, it is conducive to the balanced allocation of resources.
  • Fig. 3 is a schematic diagram showing a multi-task model training task according to an exemplary embodiment.
  • the two training data include training data 1 and training data 2, wherein the training data 1 and the training data 2 are in the same training data set.
  • the model prediction of task 1 and the model prediction of task 2 are determined.
  • the model training task is assigned to the OAM and the radio access network device, and the OAM side maintains a shared model layer shared by all gNB-CUs, which is used for model training Data feature extraction, the gNB-CU side maintains a unique model layer exclusive to gNB-CU, which is used to output model results, and the output data of the shared model layer is used as the input of the unique model layer, in gNB-CU according to the local
  • the OAM updates the structural parameters of the common model layer according to the training loss value sent by the gNB-CU.
  • the collaborative training method not only increases the generalization ability of the training model, improves user service experience, ensures the effectiveness of wireless network AI analysis services, but also reduces model training costs, which is conducive to improving the efficiency of wireless network operations.
  • the wireless communication system in the embodiment of the present disclosure is a network that provides a wireless communication function.
  • Wireless communication systems can use different communication technologies, such as code division multiple access (CDMA), wideband code division multiple access (WCDMA), time division multiple access (TDMA) , frequency division multiple access (FDMA), orthogonal frequency-division multiple access (OFDMA), single carrier frequency-division multiple access (single Carrier FDMA, SC-FDMA), carrier sense Multiple Access/Conflict Avoidance (Carrier Sense Multiple Access with Collision Avoidance).
  • CDMA code division multiple access
  • WCDMA wideband code division multiple access
  • TDMA time division multiple access
  • FDMA frequency division multiple access
  • OFDMA orthogonal frequency-division multiple access
  • single Carrier FDMA single Carrier FDMA
  • SC-FDMA carrier sense Multiple Access/Conflict Avoidance
  • Carrier Sense Multiple Access with Collision Avoidance Carrier Sense Multiple Access with Collision Avoidance
  • the network can be divided into 2G (English: generation) network, 3G network, 4G network or future evolution network, such as 5G network, 5G network can also be called a new wireless network ( New Radio, NR).
  • 2G International: generation
  • 3G network 4G network or future evolution network, such as 5G network
  • 5G network can also be called a new wireless network ( New Radio, NR).
  • New Radio New Radio
  • the present disclosure sometimes simply refers to a wireless communication network as a network.
  • the wireless access network device may be: a base station, an evolved base station (evolved node B, base station), a home base station, an access point (access point, AP) in a wireless fidelity (wireless fidelity, WIFI) system, a wireless relay Node, wireless backhaul node, transmission point (transmission point, TP) or transmission and reception point (transmission and reception point, TRP), etc., can also be gNB in the NR system, or it can also be a component or a part of equipment that constitutes a base station Wait.
  • the network device may also be a vehicle-mounted device.
  • V2X vehicle-to-everything
  • the network device may also be a vehicle-mounted device. It should be understood that in the embodiments of the present disclosure, no limitation is imposed on the specific technology and specific device form adopted by the network device.
  • terminals involved in this disclosure can also be referred to as terminal equipment, user equipment (User Equipment, UE), mobile station (Mobile Station, MS), mobile terminal (Mobile Terminal, MT), etc.
  • a device providing voice and/or data connectivity for example, a terminal may be a handheld device with a wireless connection function, a vehicle-mounted device, and the like.
  • examples of some terminals are: smart phones (Mobile Phone), pocket computers (Pocket Personal Computer, PPC), handheld computers, personal digital assistants (Personal Digital Assistant, PDA), notebook computers, tablet computers, wearable devices, or Vehicle equipment, etc.
  • V2X vehicle-to-everything
  • the terminal device may also be a vehicle-mounted device. It should be understood that the embodiment of the present disclosure does not limit the specific technology and specific device form adopted by the terminal.
  • Fig. 4 is a schematic diagram of a system architecture of a model training method according to an exemplary embodiment.
  • the system includes a terminal, radio access network equipment (wherein the radio access network equipment includes gNB-DU and gNB-CU) and OAM.
  • the terminal accesses the gNB-DU through the wireless channel, multiple gNB-DUs access the gNB-CU through the F1 interface, and the gNB-CUs are connected through the Xn interface.
  • OAM is mainly responsible for the work of model training functional units and data collection/preparation functional units in the wireless network architecture supporting AI, and is mainly responsible for sharing model training and data collection.
  • gNB-CU mainly undertakes the work of model training functional unit, data collection/preparation functional unit and model reasoning functional unit in the wireless network architecture supporting AI, and is mainly responsible for unique model training, model reasoning and data collection.
  • the gNB-DU is mainly responsible for the work of the data collection/preparation functional unit in the wireless network architecture supporting AI, and is mainly responsible for the work of data collection.
  • the terminal is mainly responsible for the work of the action execution function unit in the wireless network architecture supporting AI, and is mainly responsible for the work of policy execution and performance feedback.
  • Fig. 5 is a flow chart of a model training method according to an exemplary embodiment. As shown in Figure 5, the model training method is used in OAM, including the following steps.
  • step S11 group the multiple radio access network devices that send the model subscription request to obtain at least one radio access network device group.
  • the radio access network device group includes a first quantity of radio access network devices.
  • the OAM receives multiple model subscription requests sent by multiple radio access network devices, and determines information such as terminal identifiers, model request types, and access locations included in each model subscription request.
  • the terminal identity is a Globally Unique Temporary UE Identity (Globally Unique Temporary UE Identity, GUTI).
  • the model request type is represented by the analysis ID, such as the load forecasting analysis service.
  • the access location information mainly includes the gNB-CU and gNB-DU information currently accessed by the terminal.
  • the OAM analyzes the similarity of the models requested in the model request, and groups the model subscription requests sent by different radio access network devices according to the similarity.
  • the analysis subscription request information from gNB-CU(n1) to gNB-CU(n2) shows that the terminal requests the model for load forecasting
  • the analysis subscription request information from gNB-CU(n3) to gNB-CU(n4) shows the terminal request
  • gNB-CU(n1) to gNB-CU(n2) can be divided into a group for model training
  • gNB-CU(n3) to gNB- CU(n4) is divided into another group for model training.
  • step S12 a first quantity model training structure corresponding to the first quantity of radio access network devices is determined, and based on the first quantity model training structure, a first quantity specific model layer is determined.
  • the OAM determines the model training structure for each radio access network device group, that is, determines the corresponding first number of model training structures for the first number of radio access network devices in each radio access network device group structure.
  • the first quantity model training structure is divided into a shared model layer and a unique model layer. Determine the unique model layer corresponding to each model training structure.
  • step S13 the structural parameters of the first number of unique model layers are sent to the first number of radio access network devices.
  • the determined unique model layer corresponding to each model training structure is sent to the corresponding radio access network device.
  • OAM sends the structural parameter information of its unique model layer to each gNB-CU according to the mapping table from each gNB-CU to each connection mode; each gNB-CU receives the model information, Use the unique model layer as a local model for model training and updating.
  • part of the model training work can be transferred to wireless access network devices, which can reduce the amount of uploaded data, facilitate balanced allocation of resources, and reduce data security risks.
  • Fig. 6 is a flow chart of a model training method according to an exemplary embodiment. As shown in Figure 6, the model training method is used in OAM, including the following steps.
  • step S21 determine the first number of model subscription requests sent by the first number of radio access network devices, and determine the model training task characteristics of the first number of model subscription requests.
  • the model training task characteristic is used to indicate the number of model layers and nodes.
  • the model training tasks of the model subscription requests are different, and the number of layers and nodes of the models corresponding to the request training are also different, and the model training task characteristics of each model subscription request can be determined based on the model training tasks.
  • the number of input layer nodes is set to M, which represents the number of training data in the input model at one time.
  • the number of nodes in the output layer is set to N, and N depends on the number of gNB-CUs and the characteristics of the training task. For example, each gNB-CU in the prediction task (regression task) corresponds to one node, and each gNB-CU in the decision task corresponds to multiple node.
  • the number of hidden layers is set to S, and the number of nodes in each hidden layer is set to L. The number of hidden layers needs to consider factors such as model size and model generalization ability. Determining model training task characteristics from ears.
  • a first number of model training structures is determined based on the number of model layers and the number of nodes indicated by the characteristics of the model training task.
  • the model training structure includes the connection mode between the model layers and the connection mode between the corresponding layers.
  • the hidden layer and the input layer may be fully connected, and the activation function may use the ReLU function.
  • the hidden layer can be fully connected, and the activation function can use the ReLU function.
  • the hidden layer and the output layer can be partially connected, and the activation function can use the Softmax function or the Sigmoid function.
  • the prediction task (regression task) using the mean square error (MSE) loss function, mean absolute error (MAE) loss function, Huber loss function, etc.
  • MSE mean square error
  • MAE mean absolute error
  • Huber loss function etc.
  • the decision task (Classification task) Use cross-entropy loss function, Hinge loss function, logarithmic loss function, etc.
  • the process of determining the hyperparameters of the corresponding network model can refer to the setting of learning rounds as T times.
  • the setting of learning rounds needs to measure the impact of model training speed, training cost and model training accuracy.
  • the learning rate is set to ⁇ and ⁇ .
  • the method of weight initialization selects random weight initialization.
  • Fig. 7 is a schematic flowchart of a training model structure of a model training method according to an exemplary embodiment. As shown in Figure 7, the following steps are included:
  • step S211 firstly, the OAM determines the model structure of the training model according to information such as the number of gNB-CUs in the gNB-CU group and the characteristics of the training task.
  • step S212 the OAM divides the training model into a shared model layer and a unique model layer, wherein the shared model layer is commonly used by all gNB-CUs, and the unique model layer is used individually by each gNB-CU.
  • step S213 the OAM determines the connection mode between the shared model layer and the unique model layer, and initializes model parameters.
  • step S214 the OAM sends the structure and parameters of the unique model layer of each gNB-CU to the corresponding gNB-CU.
  • Fig. 8 is a flowchart showing a model training method according to an exemplary embodiment. As shown in Figure 8, the model training method is used in OAM, including the following steps.
  • step S31 the output layer of the first quantity model training structure corresponding to the first quantity of radio access network devices is determined as the first quantity specific model layer.
  • the output layer of the model training structure corresponding to each radio access network device in the first data radio access network device is determined, and the output layer is determined as a unique model layer to obtain a first number of unique models Floor.
  • each unique model layer is used independently by the corresponding radio access network device, and is used to output the final classification or regression result.
  • Fig. 9 is a flowchart showing a model training method according to an exemplary embodiment. As shown in Figure 9, the model training method is used in OAM, including the following steps.
  • step S41 the input layer and the hidden layer of the model training structure corresponding to the first quantity of wireless access network devices are determined as the shared model layer, and the data of multiple wireless access network devices is obtained, and the data is added to each The data identifier corresponding to the radio access network device.
  • the input layer and the hidden layer of the model training structure corresponding to each radio access network device in the first data radio access network device are determined, and the input layer and the hidden layer are determined as the shared model layer.
  • the shared model layer is commonly used by all gNB-CUs to extract feature information of input data.
  • the OAM can also request the wireless access network device to obtain model training data, perform data processing, and identify each piece of training data, identifying the wireless access network device to which the data belongs and the data ID, etc. information. Further, the OAM sends a model training data request to each managed radio access network device. After receiving the model training data request, the wireless access network device sends the model training data request to each connected wireless access network device. Each wireless access network device sends a model training data request to a terminal connected to the wireless access network device, and after receiving the model training data request, the terminal collects terminal data and sends it to the wireless access network device.
  • the wireless access network device summarizes the terminal training data, collects the data of the wireless access network device, and sends it to the wireless access network device connected to it.
  • the wireless access network device summarizes the data of the wireless access network device, collects the data of the wireless access network device, and sends it to the OAM.
  • the wireless access network device information can be identified by 0-1 encoding. Assuming that the number of wireless access network devices is N, N bits are used to record the wireless access network device information. If each bit is 0, it represents the data Does not belong to the corresponding gNB-CU. If it is 1, it means that the data belongs to the corresponding radio access network device; the data ID information needs to be consistent with the data ID information at the radio access network device.
  • step S42 classify all data with data identifiers to obtain model training data and model label values.
  • OAM performs data processing on the model training data, such as data denoising, normalization, etc., to obtain data for model training, including model training data of the shared model layer and model label values of the unique model layer .
  • the amount of data corresponding to the training task can be expanded to realize partial data sharing. And sharing data weakens the network capability to a certain extent and reduces the risk of overfitting.
  • Fig. 10 is a schematic flowchart of collecting and identifying model training data for a training model structure according to an exemplary embodiment. As shown in Figure 10, taking the radio access network device as gNB-CU as an example, the following steps are included:
  • step S421 the OAM sends a data collection request to each gNB-CU.
  • Step S422 each gNB-CU collects terminal data and sends it to the OAM.
  • step S423 the OAM summarizes the data sent by each gNB-CU to form model training data.
  • step S424 the OAM performs data processing on the model training data, such as data denoising and normalization.
  • step S425 the OAM performs data identification on the model training data, identifies the gNB-CU information to which each data record belongs, and the corresponding ID information.
  • step S43 the model training data is used as the first input data and input to the shared model layer to obtain the first output data output by the shared model layer.
  • OAM uses the model training data as the input of the shared model layer, and OAM inputs the model training data to the shared model layer in a serial manner, and the last layer of the shared model layer has L nodes, so each A piece of model training data i corresponds to a set of output results All the output results of the shared model layer are obtained, that is, the first output data output by the shared model layer.
  • the output results of the shared model layer are used as the first output data for convenience of description.
  • step S44 the model label value and the first output data are sent to multiple radio access network devices.
  • the OAM after obtaining all the output results of the model training data (that is, the first output data), the OAM sends the first output data and the identification information model tag value of the model training data to each wireless access network device .
  • the model training data provided in the present disclosure can make multi-model training tasks to be performed collaboratively, and noise will be added to each other, thereby improving the generalization ability of the model.
  • Fig. 11 is a flowchart showing a model training method according to an exemplary embodiment. As shown in Figure 11, the model training method is used in OAM, including the following steps.
  • step S51 in response to receiving the training loss values sent by multiple wireless access network devices, the structural parameters of the shared model layer are updated based on the training loss values.
  • the OAM after receiving the training loss values sent by multiple radio access network devices, the OAM updates the shared model layer structure parameters according to each loss value.
  • Fig. 12 is a flowchart of a model training method according to an exemplary embodiment. As shown in Figure 12, the model training method is used in OAM, including the following steps.
  • step S61 the training loss value is weighted to obtain a weighted loss value.
  • the OAM weights the training loss value based on factors such as the data volume and learning effect of each gNB-CU.
  • the weighting of the training loss value can refer to the following formula:
  • the calculation of w k includes two aspects, one is the proportion of the training data volume of each gNB-CU to the total data volume, and the other is the influence of the learning effect, such as the accuracy of the training model and the difficulty of the learning task degree etc.
  • step S62 the current model parameters and model learning rate of the shared model layer are determined.
  • step S63 based on the weighted loss value, the model parameters and the model learning rate, the update parameters of the shared model layer are determined, and the structural parameters of the shared model layer are updated based on the update parameters.
  • the OAM determines to use the weighted training loss value, the model updating method, and the selected structural parameters to update the structural parameters of the shared model layer.
  • the SGD algorithm is used to update the parameters of the shared model layer, as shown in the following formula:
  • b t represents the shared model layer structure parameters to be updated in round t
  • loss t represents the weighted training loss value calculated in round t
  • ⁇ t represents the learning rate in round t.
  • Fig. 13 is a flowchart of a model training method according to an exemplary embodiment. As shown in Figure 13, the model training method is used in OAM, including the following steps.
  • step S71 in response to updating the structural parameters of the shared model layer for the Tth time, it is determined that the training of the shared model layer is completed, and the model structural parameters of the shared model layer after the T-th update are sent to multiple wireless access network device groups Each wireless access network device.
  • T is the preset number of times for updating the shared model layer and the unique model layer
  • the structural parameters of the shared model layer are used by the radio access network device to synthesize the model subscribed by the radio access network device.
  • the OAM after completing the update of the structural parameters of the shared model layer for the T time, the OAM sends the updated structural parameters of the shared model layer for the T time to each radio access network device.
  • the wireless access network device receives the model information, because the wireless access network device saves the connection mode of the shared model layer and the unique model layer, the wireless access network device can splice the two models according to a specific connection mode, Integrate into a complete model for model inference.
  • Fig. 14 is a flow chart showing a model training method according to an exemplary embodiment. As shown in Figure 14, the model training method is used in OAM, including the following steps.
  • step S81 the type of the subscription model included in each model subscription request is determined.
  • the OAM determines the type of the requested model training task included in the model subscription request sent by each radio access network device, such as a load forecasting model training task, a network decision training task, and the like.
  • step S82 the model subscription requests are grouped based on types to obtain a first group of model subscription requests.
  • OAM groups the received model subscription requests, determines different groups of model subscription requests, and further obtains the first group of models Subscribe request.
  • step S83 the radio access network devices are grouped to obtain a first number of radio access network device groups.
  • the OAM groups corresponding radio access network devices according to the first group of model subscription requests to obtain the first group of radio access network device groups.
  • OAM adopts a training method of cooperative training. Under the cooperative training method, the more similar the training tasks of the wireless access network devices participating in the training, the better the training effect. Therefore, the wireless access network devices are grouped for training.
  • Fig. 15 is a flowchart of a model training method according to an exemplary embodiment. As shown in Figure 15, the model training method is used in OAM, including the following steps.
  • step S91 in response to the presence of newly added wireless access network equipment, and the newly added wireless access network equipment meets the model training conditions, the structural parameters of the unique model layer corresponding to the newly added wireless access network equipment are sent to to the newly added wireless access network device.
  • the new radio access network device when the new radio access network device requests to join the radio access network device group, the new radio access network device first sends a request to join the current gNB-CU group to the OAM. After receiving the request, OAM judges whether the new wireless access network device meets the conditions for joining the current wireless access network device group. If the conditions are met, it can join the current wireless access network device group to participate in model training. Cannot join the current wireless access network device group. If the conditions are met, the OAM updates the information of the radio access network device group, and adds the new radio access network device to the list of radio access network devices participating in model training.
  • the OAM updates the training model structure according to information such as the training task characteristics of the new radio access network device, and sends the unique model layer structure and parameters of the new radio access network device to the new radio access network device.
  • the OAM sends the output result of the shared model layer to the new wireless access network device, and the new wireless access network device updates the structural parameters of the unique model layer to perform model training.
  • Fig. 16 is a flow chart of adding a radio access network device according to a model training method according to an exemplary embodiment. As shown in Figure 16, the following steps are included:
  • step S911 the new gNB-CU first sends a request to OAM to join the current gNB-CU group.
  • Step S912 after receiving the request, the OAM first judges whether the new gNB-CU meets the conditions for joining the current gNB-CU group. If the conditions are met, it can join the current gNB-CU group to participate in model training. If the conditions are not met, it cannot Join the current gNB-CU group.
  • Step S913 if the condition is satisfied, the OAM updates the information of the gNB-CU group, and adds the new gNB-CU to the gNB-CU list participating in the model training.
  • step S914 the OAM updates the training model structure according to information such as the training task characteristics of the new gNB-CU, and sends the unique model layer structure and parameters of the new gNB-CU to the new gNB-CU.
  • step S915 the OAM sends the output result of the shared model layer to the new gNB-CU, and the new gNB-CU updates the parameters of the unique model layer to perform model training.
  • a new terminal sends an analysis subscription request to the gNB-CU, and then a new gNB-CU sends a model subscription request to the OAM.
  • the OAM first judges whether the gNB-CU has the conditions to join the current gNB-CU group.
  • the method for judging whether the new gNB-CU has the conditions to join the current gNB-CU group is to compare the analysis request type in the terminal model subscription request information of the new gNB-CU with the terminal model subscription information of the gNB-CU group Analyze the similarity of the request type in the request information.
  • the new gNB-CU meets the conditions and can join the current gNB-CU group to participate in model training. If the similarity is low, the new gNB-CU does not meet the conditions. You cannot join the current gNB-CU group to participate in model training. For example, if a new gNB-CU requests a training model to perform a prediction task, and the training model of this gNB-CU group is used for a decision-making task, the similarity between the two is low, and this gNB-CU cannot be added to the current gNB-CU group.
  • the OAM will add it to the list for participating in model training, update the information of the gNB-CU group, and start sending data information to it.
  • OAM modifies the structure of the unique model layer without changing the shared model layer, including adding branches, increasing the number of output layer nodes, changing the connection mode with the shared model layer, etc.
  • the structure is updated, and the newly added unique model layer structure and parameters are sent to the new gNB-CU as the unique model layer of the new gNB-CU.
  • Fig. 17 is a flowchart of a model training method according to an exemplary embodiment. As shown in Figure 17, the model training method is used in OAM, including the following steps.
  • step S101 in response to exiting radio access network devices, a first quantity model training structure is re-determined.
  • the radio access network device when the radio access network device requests to withdraw from the radio access network device group, the radio access network device first sends a request to the OAM to withdraw from the current radio access network device group. After receiving the exit request, OAM deletes the relevant information of the radio access network device in the list of radio access network devices participating in model training, and no longer sends data to the radio access network device; OAM deletes the radio access network device The unique model layer of the update training model structure. The wireless access network device no longer participates in the model training process of the wireless access network device group. If the wireless access network device has not completed the current round of model training, the wireless access network device continues to complete the current round of model training , but no more parameter uploads.
  • Fig. 18 is a flow chart of a radio access network device exiting a model training method according to an exemplary embodiment. As shown in Figure 18, it includes the following steps:
  • step S1011 the gNB-CU sends a request to the OAM to withdraw from the current gNB-CU group.
  • step S1012 after receiving the exit request, the OAM deletes the relevant information of the gNB-CU from the list of gNB-CUs participating in model training, and no longer sends data to the gNB-CU.
  • step S1013 the OAM deletes the specific model layer of the gNB-CU, and updates the training model structure.
  • Step S1014 the gNB-CU no longer participates in the model training process of the gNB-CU group.
  • the gNB-CU during the training process of the current gNB-CU group, there is a terminal request to cancel the analysis subscription to the gNB-CU, and then the gNB-CU sends a request to cancel the model subscription to the OAM, requesting to exit the current gNB-CU CU group.
  • the OAM deletes the gNB-CU from the list participating in model training, updates the information of the gNB-CU group, and no longer sends data information to it.
  • OAM modifies the structure of the unique model layer without changing the shared model layer, including deleting branches, reducing the number of output layer nodes, changing the connection mode with the shared model layer, etc.
  • the structure is updated, and the unique model layer structure of the gNB-CU is deleted.
  • the gNB-CU no longer participates in the model training process of the gNB-CU group. If the gNB-CU has not completed the current round of model training, the gNB-CU will continue to complete the current round of model training, but no longer upload parameters.
  • the embodiment of the present disclosure also provides a model training method.
  • Fig. 19 is a flow chart showing a model training method according to an exemplary embodiment. As shown in Figure 19, the model training method is used in radio access network equipment, including the following steps.
  • step S111 the structural parameters of the unique model layer sent by the OAM are received.
  • the wireless access network device receives the unique model layer sent by OAM.
  • the unique model layer is determined by the OAM to divide the first number of model training structures; the first number of model training structures is determined by the OAM based on the model subscription requests of the first number of radio access network devices included in the radio access network device group.
  • OAM sends the structural parameter information of its unique model layer to each gNB-CU according to the mapping table from each gNB-CU to each connection mode; each gNB-CU receives the model layer Information, using the unique model layer as a local model for model training and updating.
  • part of the model training work can be transferred to wireless access network devices, which can reduce the amount of uploaded data, facilitate balanced allocation of resources, and reduce data security risks.
  • Fig. 20 is a flowchart of a model training method according to an exemplary embodiment. As shown in Figure 20, the model training method is used in radio access network equipment, including the following steps.
  • step S121 the model tag value and the first output data sent by OAM are received.
  • the wireless access network device receives the first output data and the model tag value sent by OAM, and each wireless access network device determines that it belongs to the wireless access network according to the identification information of the wireless access network device of the model tag value.
  • the model label values of network devices and filter out the first output data output by the shared model layer of these model label values.
  • step S122 the first output data is input to the unique model layer as an input of the unique model layer to obtain second output data output by the unique model layer.
  • the wireless access network device filters the first output data output by the shared model layer corresponding to itself, and inputs it into the unique model layer received by the wireless access network device, and obtains the second output data of the unique model layer. Output Data.
  • each wireless access network device stores the structure and parameter information of the unique model layer sent by OAM, including the connection mode of the unique model layer and the shared model layer, and each wireless access network device In the connection mode, the output results of the shared model layer can be input into the unique model layer.
  • the radio access network device serially inputs the output result of the shared model layer that belongs to the radio access network device to the unique model layer to obtain the output result of the unique model layer, and the number of nodes in the output layer is N, so each piece of model training data i corresponds to a set of output results
  • step S123 a training loss value is determined based on the model label value and the second output data, and the training loss value is sent to the OAM.
  • the radio access network device determines a model label value corresponding to itself, and determines a training loss value according to the model label value and the second output data.
  • the determination of the training loss value can refer to the following formula:
  • loss is the training loss value
  • I is the amount of model training data belonging to this gNB-CU
  • y i is the output result of data i passing through the unique model layer
  • an appropriate loss function may also be selected according to different tasks.
  • Fig. 21 is a flowchart of a model training method according to an exemplary embodiment. As shown in Figure 21, the model training method is used in radio access network equipment, including the following steps.
  • step S131 based on the identifier carried in the model training data, among the model tag values, a model tag value corresponding to the radio access network device is determined.
  • each wireless access network device searches in the database of the wireless access network device according to the ID information of the model training data judged to belong to itself, and obtains information such as the tag value of each model training data .
  • the radio access network device judges whether the training data belongs to the radio access network device by analyzing the N-bit radio access network device information carried by it, and then determines all the information belonging to the radio access network device. Obtain the training data of the networked device, obtain its data ID information and the output results after passing through the shared model layer.
  • step S132 calculate the second output data and the training label value, determine the training loss value, and update the structural parameters of the unique model layer based on the training loss value.
  • each gNB-CUk uses its own training loss value, model update method, and selected structural parameters, such as stochastic gradient descent algorithm (Stochastic Gradient Descent, SGD), optimization algorithm Adam, etc., for the unique model layer
  • SGD stochastic gradient descent algorithm
  • Adam optimization algorithm Adam
  • gNB-CU is taken as an example for illustration.
  • Fig. 22 is a flowchart showing a method for updating a model-specific layer parameter according to an exemplary embodiment. As shown in Figure 22, the following steps are included:
  • each gNB-CU determines which training data belongs to its own gNB-CU according to the gNB-CU identification information of the training data, and screens out the shared model layer output results of these data.
  • Step S1322 each gNB-CU obtains information such as the label value of the training data according to the identification information of the training data.
  • each gNB-CU uses the output result of the shared model layer as the input of the unique model layer to obtain the output result of the unique model layer.
  • Step S1324 each gNB-CU calculates the training loss value according to the output result of the unique model layer and the label value information.
  • Step S1325 each gNB-CU updates its unique model layer parameters according to the training loss value.
  • Fig. 23 is a flow chart of a model training method according to an exemplary embodiment. As shown in Figure 23, the model training method is used in radio access network equipment, including the following steps.
  • step S141 the structural parameters of the shared model layer sent by the OAM are received.
  • step S142 the structural parameters of the subscription model are determined based on the structural parameters of the model sharing layer and the structural parameters of the unique model layer after the T-th update.
  • T is the preset number of times to update the shared model layer and the unique model layer.
  • the OAM sends the model parameters of the shared model layer to each radio access network device in the radio access network device group, and after the radio access network device receives the structural parameters of the shared model layer, the radio access network Based on the connection method between the saved shared model layer and the unique model layer, the network access device splices the structural parameters of the two models according to the specific connection method, and integrates them into a complete model, and obtains the structural parameters of the model, which can be used in the model reasoning.
  • the complete model may be used for model reasoning, and the model reasoning process includes:
  • the wireless access network equipment collects model reasoning data, performs model reasoning based on the reasoning model, sends the reasoning result to the terminal, and feeds back the model reasoning result to the OAM.
  • the radio access network device sends a model training data request to each connected radio access network device.
  • Each radio access network device sends a model training data request to all terminals accessing the radio access network device.
  • the terminal collects the terminal data and sends it to the wireless access network device.
  • the wireless access network device summarizes the terminal data, collects the data of the wireless access network device, and sends it to the wireless access network device connected to it.
  • the wireless access network device summarizes the data of the wireless access network device, and collects the data of the wireless access network device to form model reasoning data.
  • the wireless access network device After the wireless access network device collects the model reasoning data, it performs model reasoning based on the integrated reasoning model, and sends the reasoning result to the terminal.
  • the radio access network device sends the reasoning result to the radio access network device accessed by the terminal.
  • the radio access network device sends the inference result to the terminal.
  • the radio access network device After the radio access network device completes the model reasoning, it feeds back the reasoning result to the OAM.
  • the reasoning result that the radio access network device needs to feed back is information such as accuracy of model reasoning.
  • the terminal executes the network optimization strategy according to the model reasoning result, collects the network performance data and feeds it back to the OAM for model training.
  • the terminal executes corresponding network optimization strategies (such as cell handover, cell activation, etc.) according to the model reasoning results (such as prediction results, decision results, etc.); at the same time, the terminal will collect performance data on the network side (such as measurement results, cell handover successful or failure related data, etc.), and feed back to the radio access network device, and the radio access network device feeds back the performance data to the OAM for model training.
  • network optimization strategies such as cell handover, cell activation, etc.
  • model reasoning results such as prediction results, decision results, etc.
  • the terminal will collect performance data on the network side (such as measurement results, cell handover successful or failure related data, etc.), and feed back to the radio access network device, and the radio access network device feeds back the performance data to the OAM for model training.
  • Fig. 24 is a schematic diagram of model training and model inference of a model training method according to an exemplary embodiment. As shown in Figure 24, it includes a structure of a shared model layer and a unique model layer, wherein the shared model layer includes an input layer and a hidden layer, and the unique model layer includes an output layer.
  • the shared model layer is trained on the OAM, and the unique model layer Training is performed on the network access device (such as gNB-CU), and each gNB-CU only retains one branch of the unique model layer.
  • the network access device such as gNB-CU
  • OAM obtains the output result of the shared model layer and sends it to each gNB-CU; each gNB-CU then obtains the output result of its unique model layer according to the output result of the shared model layer, and calculates the training loss value, and updates The model parameters of the unique model layer; each gNB-CU sends the training loss value to the OAM, and the OAM updates the model parameters of the shared model layer to continue training.
  • OAM sends the model information of the trained shared model layer to each gNB-CU; each gNB-CU integrates the shared model layer with the local specific model layer to form a complete inference model; each gNB-CU uses the inference model to perform model reasoning.
  • Fig. 25 is a flow chart showing a model training method and a model reasoning method according to an exemplary embodiment. As shown in Figure 25, the following steps are included:
  • step S151 the terminal initiates an analysis subscription request to the gNB-CU to which it belongs, and each gNB-CU acquires information such as task characteristics of local training according to the request of its own terminal.
  • Step S152 each gNB-CU sends a model subscription request to the OAM.
  • step S153 the OAM summarizes the model subscription requests of each gNB-CU, and groups the gNB-CUs according to the similarity of the training tasks to obtain different gNB-CU groups.
  • Step S154 for each gNB-CU group, OAM determines the appropriate training model structure, divides the training model into a shared model layer and a unique model layer, initializes model parameters, and sends the unique model layer structure parameters of each gNB-CU to Corresponding gNB-CU.
  • step S155 the OAM collects model training data, performs data processing, and identifies each piece of training data, identifying the gNB-CU to which the data belongs, the data ID and other information.
  • step S156 the OAM takes the model training data as input, obtains the output result of the shared model layer, and sends the output result and identification information corresponding to the training data to each gNB-CU.
  • Step S157 each gNB-CU screens according to the identification information of the training data and obtains the model label value of the training data, uses the output result of the shared model layer as the input of the unique model layer, obtains the output result and calculates the training loss value, and calculates the training loss value for the unique model Layer parameters are updated.
  • each gNB-CU sends a training loss value to the OAM, and the OAM updates the structural parameters of the shared model layer according to each loss value.
  • step S159 when the gNB-CU needs to join or quit the gNB-CU group, the gNB-CU sends a join or quit request to the OAM, and completes the process of joining or quitting the gNB-CU group.
  • step S160 after the model training is completed, the OAM sends the shared model layer structure parameters to each gNB-CU, and the gNB-CU receives and integrates the model structure parameters to form a complete reasoning model.
  • step S161 the gNB-CU collects model inference data, performs model inference based on the inference model, sends the inference result to the terminal, and feeds back the model inference result to the OAM.
  • step S162 the terminal executes a network optimization strategy according to the model inference result, collects network performance data, and feeds back to the gNB-CU and OAM for model training.
  • Fig. 26 is a schematic diagram of a model training protocol and interface of a model training method according to an exemplary embodiment. As shown in FIG. 26 , it mainly involves the terminal, the radio access network device (such as gNB-DU) accessed by the terminal, the radio access network device (such as gNB-CU) accessed by the terminal and OAM provided by the embodiment of the present disclosure. details as follows:
  • the terminal sends analysis subscription request signaling to the gNB-DU.
  • the gNB-DU receives the analysis subscription request signaling sent by the terminal, and sends the analysis subscription request signaling to the gNB-CU.
  • the gNB-CU receives the analysis subscription request signaling sent by the gNB-DU, and forms a model subscription request signaling.
  • the gNB-CU sends the model subscription request signaling to the OAM.
  • the OAM receives the model subscription request signaling, obtains the information contained in the signaling, and groups the gNB-CUs. 5.
  • the OAM sends the grouping information to each gNB-CU in the gNB-CU group. 6.
  • OAM determines the training model structure and divides the model into a shared model layer and a unique model layer. 7. The OAM sends the model information of the unique model layer to each gNB-CU in the gNB-CU group. 8. Each gNB-CU receives the model information of the unique model layer for model training. 9. OAM collects model training data, and processes and identifies the data after data collection. 10. OAM uses the model training data as the input of the shared model layer to obtain the output result. 11. The OAM sends the output results of the shared model layer and the identification information of the training data to each gNB-CU. 12.
  • Each gNB-CU screens out the output results corresponding to the training data belonging to the gNB-CU according to the identification information of the training data, and obtains the label values of these training data. 13.
  • Each gNB-CU uses the output result of the screened shared model layer as the input of the unique model layer, obtains the output result of the unique model layer, calculates the training loss value, and updates the parameters of the unique model layer. 14.
  • Each gNB-CU sends the calculated training loss value to the OAM.
  • OAM receives and summarizes each training loss value, and updates the parameters of the shared model layer.
  • Fig. 27 is a schematic diagram of the protocol and interface of model inference in a model training method according to an exemplary embodiment. As shown in FIG. 27 , it mainly involves the terminal, the radio access network device (such as gNB-DU) accessed by the terminal, the radio access network device (such as gNB-CU) accessed by the terminal and OAM provided by the embodiment of the present invention. details as follows:
  • Each gNB-CU sends the model request signaling to the OAM.
  • the OAM receives the model request signaling and prepares the model information of the shared model layer.
  • the OAM sends the model information of the shared model layer to each gNB-CU.
  • Each gNB-CU receives the shared model layer information and integrates it with the gNB-CU's local specific model layer to form a complete reasoning model.
  • Each gNB-CU collects model reasoning data, performs model reasoning based on the reasoning model, and obtains reasoning results. 6a.
  • Each gNB-CU sends the model inference result to the connected gNB-DU.
  • 6b Each gNB-DU sends the model inference result to the connected terminal. 6c.
  • Each gNB-CU sends the model reasoning feedback result to the OAM. 7.
  • the terminal executes network optimization strategies and collects network performance data according to the model reasoning results.
  • the terminal feeds back the network performance data to the connected gNB-DU.
  • the gNB-DU feeds back network performance data to the connected gNB-CU.
  • the gNB-CU feeds back the network performance data to the OAM.
  • gNB-CU and OAM receive network performance feedback data for model training.
  • Fig. 28 is a schematic diagram of a protocol and an interface for collecting model training data in a model training method according to an exemplary embodiment. As shown in Figure 28, it mainly involves the terminal, the radio access network device (such as gNB-DU) accessed by the terminal, the radio access network device (such as gNB-CU) accessed by the terminal and OAM provided by the embodiment of the present invention.
  • the radio access network device such as gNB-DU
  • the radio access network device such as gNB-CU
  • OAM provided by the embodiment of the present invention.
  • the OAM sends the model training data request signaling to each gNB-CU. 1b.
  • Each gNB-CU sends model training data request signaling to the connected gNB-DU.
  • Each gNB-DU sends model training data request signaling to the connected terminal. 2.
  • the terminal receives the model training data request and prepares the terminal training data. 3.
  • the terminal sends the training data to the connected gNB-DU. 4.
  • Each gNB-DU receives terminal training data and collects the data of its own gNB-DU to form gNB-DU training data. 5.
  • Each gNB-DU sends training data to the connected gNB-CU. 6.
  • Each gNB-CU receives gNB-DU training data, and collects the data of its own gNB-CU to form gNB-CU training data. 7. Each gNB-CU sends the training data to the OAM. 8. OAM receives gNB-CU training data and collects OAM local data to form model training data.
  • Fig. 29 is a schematic diagram of a protocol and interface for model inference data collection in a model training method according to an exemplary embodiment. As shown in Figure 29, it mainly involves the terminal, the radio access network device (such as gNB-DU) accessed by the terminal, the radio access network device (such as gNB-CU) accessed by the terminal and OAM provided by the embodiment of the present invention.
  • the radio access network device such as gNB-DU
  • the radio access network device such as gNB-CU
  • OAM provided by the embodiment of the present invention.
  • the current gNB-CU sends the model inference data request signaling to the connected gNB-DU.
  • Each gNB-DU sends the model reasoning data request signaling to the connected terminal.
  • the terminal receives the model inference data request and prepares the terminal inference data.
  • the terminal sends the reasoning data to the connected gNB-DU.
  • Each gNB-DU receives terminal reasoning data and collects the data of its own gNB-DU to form gNB-DU reasoning data.
  • Each gNB-DU sends inference data to the connected gNB-CU. 6.
  • the current gNB-CU receives gNB-DU inference data, and collects the data of this gNB-CU to form model inference data.
  • the embodiment of the present disclosure also provides a model training device.
  • the model training device provided by the embodiments of the present disclosure includes corresponding hardware structures and/or software modules for performing various functions.
  • the embodiments of the present disclosure can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed by hardware or computer software drives hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the technical solutions of the embodiments of the present disclosure.
  • Fig. 30 is a block diagram of a model training device according to an exemplary embodiment.
  • the apparatus 100 is applied to an OAM entity, and includes a grouping module 101 , a determining module 102 and a sending module 103 .
  • the grouping module 101 is configured to group multiple radio access network devices that send model subscription requests to obtain at least one radio access network device group, where the radio access network device group includes a first number of radio access network devices.
  • the determining module 102 is configured to determine a first quantity model training structure corresponding to the first quantity of radio access network devices, and determine a first quantity of unique model layers based on the first quantity model training structure.
  • the sending module 103 is configured to send the structural parameters of the first number of unique model layers to the first number of radio access network devices.
  • the determining module 102 is configured to determine the first quantity of model subscription requests sent by the first quantity of radio access network devices, and determine the model training task characteristics of the first quantity of model subscription requests, and the model training task characteristics are used Used to indicate the number of model layers and nodes. Based on the number of model layers and the number of nodes indicated by the characteristics of the model training task, a first number of model training structures is determined.
  • the determining module 102 is configured to determine the first quantity of model training structure output layers corresponding to the first quantity of radio access network devices as the first quantity of unique model layers.
  • the determination module 102 is further configured to input the model training structure corresponding to the first quantity of wireless access network devices into the layer and the hidden layer, determine it as a shared model layer, and obtain multiple wireless access network devices data, adding a data identifier corresponding to each wireless access network device to the data. All data with data identifiers are classified and processed to obtain model training data and model label values.
  • the model training data is used as the first input data and input to the shared model layer to obtain the first output data output by the shared model layer. Send the model tag value and the first output data to multiple radio access network devices.
  • the device further includes: an update module 104 .
  • the update module 104 is configured to update the structural parameters of the shared model layer based on the training loss values in response to receiving the training loss values sent by multiple radio access network devices.
  • the update module 104 is configured to weight the training loss value to obtain a weighted loss value. Determine the current model parameters and model learning rate for the shared model layer. Based on the weighted loss value, the model parameters and the model learning rate, the update parameters of the shared model layer are determined, and the structural parameters of the shared model layer are updated based on the update parameters.
  • the update module 104 is further configured to determine that the training of the shared model layer is completed in response to updating the structural parameters of the shared model layer for the Tth time, and send the model structural parameters of the shared model layer after the Tth update to to each radio access network device in the plurality of radio access network device groups.
  • T is the preset number of times for updating the shared model layer and the unique model layer
  • the structural parameters of the shared model layer are used by the radio access network device to synthesize the model subscribed by the radio access network device.
  • the grouping module 101 is configured to determine the type of the subscription model included in each model subscription request. Based on the type, the model subscription requests are grouped to obtain the first group of model subscription requests. The wireless access network devices are grouped to obtain the first group number of wireless access network device groups.
  • the update module 104 is also configured to respond to the presence of newly added wireless access network equipment, and the newly added wireless access network equipment meets the model training conditions, and will correspond to the newly added wireless access network equipment
  • the structural parameters of the unique model layer are sent to the newly added wireless access network device. Or, in response to exiting radio access network devices, re-determine the training structure of the first quantity model.
  • Fig. 31 is a block diagram of a model training device according to an exemplary embodiment.
  • the apparatus 200 is applied to radio access network equipment, and includes a receiving module 201 .
  • the receiving module 201 is configured to receive the structural parameters of the unique model layer sent by the OAM.
  • the unique model layer is determined by the OAM division first quantity model training structure.
  • the first quantity model training structure is determined by the OAM based on the model subscription requests of the first quantity of radio access network devices included in the radio access network device group.
  • the receiving module 201 is further configured to receive the model tag value and the first output data sent by the OAM.
  • the first output data is used as the input of the unique model layer, and is input to the unique model layer to obtain the second output data output by the unique model layer. Based on the model label value and the second output data, determine a training loss value, and send the training loss value to the OAM.
  • the device further includes: a determination module 202 .
  • the determining module 202 is configured to determine, among the model tag values, a model tag value corresponding to the radio access network device based on the identifier carried in the model training data. Calculate the second output data and the training label value, determine the training loss value, and update the structural parameters of the unique model layer based on the training loss value.
  • the receiving module 201 is further configured to receive the structural parameters of the shared model layer sent by the OAM.
  • the structural parameters of the subscription model are determined based on the structural parameters of the model sharing layer and the structural parameters of the unique model layer after the T-th update. Among them, T is the preset number of times to update the shared model layer and the unique model layer.
  • Fig. 32 is a block diagram of an apparatus 300 for model training according to an exemplary embodiment.
  • the apparatus 300 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, and the like.
  • device 300 may include one or more of the following components: processing component 302, memory 304, power component 306, multimedia component 308, audio component 310, input/output (I/O) interface 312, sensor component 314, and communication component 316 .
  • the processing component 302 generally controls the overall operations of the device 300, such as those associated with display, telephone calls, data communications, camera operations, and recording operations.
  • the processing component 302 may include one or more processors 320 to execute instructions to complete all or part of the steps of the above method. Additionally, processing component 302 may include one or more modules that facilitate interaction between processing component 302 and other components. For example, processing component 302 may include a multimedia module to facilitate interaction between multimedia component 308 and processing component 302 .
  • the memory 304 is configured to store various types of data to support operations at the device 300 . Examples of such data include instructions for any application or method operating on device 300, contact data, phonebook data, messages, pictures, videos, and the like.
  • the memory 304 can be implemented by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Magnetic Memory, Flash Memory, Magnetic or Optical Disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read-only memory
  • EPROM erasable Programmable Read Only Memory
  • PROM Programmable Read Only Memory
  • ROM Read Only Memory
  • Magnetic Memory Flash Memory
  • Magnetic or Optical Disk Magnetic Disk
  • Power component 306 provides power to various components of device 300 .
  • Power components 306 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for device 300 .
  • the multimedia component 308 includes a screen that provides an output interface between the device 300 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user.
  • the touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensor may not only sense a boundary of a touch or swipe action, but also detect duration and pressure associated with the touch or swipe action.
  • the multimedia component 308 includes a front camera and/or a rear camera. When the device 300 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera can receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capability.
  • the audio component 310 is configured to output and/or input audio signals.
  • the audio component 310 includes a microphone (MIC), which is configured to receive external audio signals when the device 300 is in operation modes, such as call mode, recording mode and voice recognition mode. Received audio signals may be further stored in memory 304 or sent via communication component 316 .
  • the audio component 310 also includes a speaker for outputting audio signals.
  • the I/O interface 312 provides an interface between the processing component 302 and a peripheral interface module, which may be a keyboard, a click wheel, a button, and the like. These buttons may include, but are not limited to: a home button, volume buttons, start button, and lock button.
  • Sensor assembly 314 includes one or more sensors for providing various aspects of status assessment for device 300 .
  • the sensor component 314 can detect the open/closed state of the device 300, the relative positioning of components, such as the display and keypad of the device 300, and the sensor component 314 can also detect a change in the position of the device 300 or a component of the device 300 , the presence or absence of user contact with the device 300 , the device 300 orientation or acceleration/deceleration and the temperature change of the device 300 .
  • the sensor assembly 314 may include a proximity sensor configured to detect the presence of nearby objects in the absence of any physical contact.
  • Sensor assembly 314 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications.
  • the sensor component 314 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor or a temperature sensor.
  • the communication component 316 is configured to facilitate wired or wireless communication between the apparatus 300 and other devices.
  • the device 300 can access wireless networks based on communication standards, such as WiFi, 2G or 3G, or a combination thereof.
  • the communication component 316 receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 316 also includes a near field communication (NFC) module to facilitate short-range communication.
  • NFC near field communication
  • the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra Wide Band (UWB) technology, Bluetooth (BT) technology and other technologies.
  • RFID Radio Frequency Identification
  • IrDA Infrared Data Association
  • UWB Ultra Wide Band
  • Bluetooth Bluetooth
  • apparatus 300 may be programmed by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable A gate array (FPGA), controller, microcontroller, microprocessor or other electronic component implementation for performing the methods described above.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA field programmable A gate array
  • controller microcontroller, microprocessor or other electronic component implementation for performing the methods described above.
  • non-transitory computer-readable storage medium including instructions, such as the memory 304 including instructions, which can be executed by the processor 320 of the device 300 to implement the above method.
  • the non-transitory computer readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, and the like.
  • Fig. 33 is a block diagram of an apparatus 400 for model training according to an exemplary embodiment.
  • the apparatus 400 may be provided as a server.
  • apparatus 400 includes processing component 422 , which further includes one or more processors, and a memory resource represented by memory 432 for storing instructions executable by processing component 422 , such as application programs.
  • the application program stored in memory 432 may include one or more modules each corresponding to a set of instructions.
  • the processing component 422 is configured to execute instructions to perform the above method.
  • Device 400 may also include a power component 426 configured to perform power management of device 400 , a wired or wireless network interface 450 configured to connect device 400 to a network, and an input-output (I/O) interface 458 .
  • the device 400 can operate based on an operating system stored in the memory 432, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.
  • “plurality” in the present disclosure refers to two or more, and other quantifiers are similar thereto.
  • “And/or” describes the association relationship of associated objects, indicating that there may be three types of relationships, for example, A and/or B may indicate: A exists alone, A and B exist simultaneously, and B exists independently.
  • the character “/” generally indicates that the contextual objects are an “or” relationship.
  • the singular forms “a”, “said” and “the” are also intended to include the plural unless the context clearly dictates otherwise.
  • first, second, etc. are used to describe various information, but the information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another, and do not imply a specific order or degree of importance. In fact, expressions such as “first” and “second” can be used interchangeably.
  • first information may also be called second information, and similarly, second information may also be called first information.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本公开是关于一种模型训练方法、模型训练装置及存储介质。其中，模型训练方法，应用于操作维护管理OAM实体，所述方法包括：对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组，所述无线接入网设备组包括第一数量无线接入网设备；确定所述第一数量无线接入网设备对应的第一数量模型训练结构，并基于所述第一数量模型训练结构，确定第一数量特有模型层；向所述第一数量无线接入网设备，发送所述第一数量特有模型层的结构参数。通过本公开可以实现可以将一部分模型训练的工作转移到无线接入网设备，有利于资源的均衡分配，降低数据安全风险。

Description

一种模型训练方法、模型训练装置及存储介质 技术领域
本公开涉及无线通信技术领域，尤其涉及一种模型训练方法、模型训练装置及存储介质。
背景技术
无线网络AI架构为实现无线人工智能提供了基础，另外，根据终端具有高速移动性的场景，为保证实现模型训练和模型推理的连贯性，终端所获得的AI分析服务的连续性，以及对无线人工智能进行移动性管理，对该无线网络AI架构进行规范和优化。在第三代合作伙伴计划(3rd Generation Partnership Project，3GPP)会议中，提出一种支持人工智能的无线网络架构，以得到大数据使能的人工智能无线网络。
支持人工智能的无线网络架构可以同时处理多个训练任务场景。但是，需要对每一个训练任务单独训练模型，造成相对较大的训练开销，以及降低数据安全风险。
发明内容
为克服相关技术中存在的问题，本公开提供一种模型训练方法、模型训练装置及存储介质。
根据本公开实施例的第一方面，提供一种模型训练方法，应用于操作维护管理OAM实体，所述方法包括：
对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组，所述无线接入网设备组包括第一数量无线接入网设备；确定所述第一数量无线接入网设备对应的第一数量模型训练结构，并基于所述第一数量模型训练结构，确定第一数量特有模型层；向所述第一数量无线接入网设备，发送所述第一数量特有模型层的结构参数。
一种实施方式中，所述确定所述第一数量无线接入网设备对应的第一数量模型训练结构，包括：
确定所述第一数量无线接入网设备发送的第一数量模型订阅请求，并确定所述第一数量模型订阅请求的模型训练任务特性，所述模型训练任务特性用于指示模型层数和节点数；基于所述模型训练任务特性所指示的模型层数和节点数，确定第一数量的模型训练结构。
一种实施方式中，所述基于所述第一数量模型训练结构，确定第一数量特有模型层，包括：
将与第一数量无线接入网设备对应的第一数量模型训练结构输出层，确定为第一数量 特有模型层。
一种实施方式中，所述方法还包括：
将与第一数量无线接入网设备对应的模型训练结构输入层和隐藏层，确定为共享模型层，并获取所述多个无线接入网设备的数据，对所述数据添加与每个无线接入网设备对应的数据标识；对带有所述数据标识的所有数据进行分类处理，得到模型训练数据和模型标签值；将所述模型训练数据作为第一输入数据，输入至所述共享模型层，得到所述共享模型层输出的第一输出数据；向所述多个无线接入网设备发送所述模型标签值和所述第一输出数据。
一种实施方式中，所述方法还包括：
响应于接收到所述多个无线接入网设备发送的训练损失值，基于所述训练损失值更新所述共享模型层的结构参数。
一种实施方式中，所述基于所述训练损失值更新所述共享模型层，包括：
对所述训练损失值进行加权，得到加权损失值；确定所述共享模型层当前的模型参数和模型学习率；基于所述加权损失值、模型参数和模型学习率，确定所述共享模型层的更新参数，并基于所述更新参数对所述共享模型层的结构参数进行更新。
一种实施方式中，所述基于所述更新参数对所述共享模型层进行更新之后，所述方法包括：
响应于第T次更新所述共享模型层的结构参数，确定所述共享模型层训练完成，将第T次更新后的所述共享模型层的模型结构参数，发送至所述多个无线接入网设备组中的每个无线接入网设备；其中，所述T为预设更新共享模型层和特有模型层的次数，所述共享模型层结构参数用于无线接入网设备合成无线接入网设备订阅的模型。
一种实施方式中，所述对发送模型订阅请求的无线接入网设备进行分组，得到至少一个无线接入网设备组，包括：
确定每个所述模型订阅请求包括的订阅模型的类型；基于所述类型，对所述模型订阅请求进行分组，得到第一组数的模型订阅请求；对无线接入网设备进行分组，得到所述第一组数的无线接入网设备组。
一种实施方式中，所述方法还包括：
响应于存在新增的无线接入网设备，且所述新增的无线接入网设备满足模型训练条件，将与新增无线接入网设备对应的特有模型层的结构参数，发送至所述新增的无线接入网设备；或，响应于存在退出的无线接入网设备，重新确定所述第一数量模型训练结构。
根据本公开实施例的第二方面，提供一种模型训练方法，应用于无线接入网设备，所 述方法包括：
接收OAM发送的特有模型层的结构参数；其中，所述特有模型层为OAM划分第一数量模型训练结构确定的；所述第一数量模型训练结构为OAM基于无线接入网设备组包括的第一数量无线接入网设备的模型订阅请求确定的。
一种实施方式中，所述方法还包括：
接收OAM发送的模型标签值和第一输出数据；将所述第一输出数据作为所述特有模型层的输入，输入至所述特有模型层，得到所述特有模型层输出的第二输出数据；基于所述模型标签值和所述第二输出数据，确定训练损失值，并向OAM发送所述训练损失值。
一种实施方式中，所述基于所述模型训练数据和所述第二输出数据，确定训练损失值，包括：
基于所述模型训练数据携带的标识，在所述模型标签值中，确定与所述无线接入网设备对应的模型标签值；运算所述第二输出数据与所述训练标签值，确定训练损失值，并基于所述训练损失值对所述特有模型层的结构参数进行更新。
一种实施方式中，所述方法还包括：
接收OAM发送的共享模型层的结构参数；基于所述模型共享层的结构参数和第T次更新后的特有模型层的结构参数确定订阅模型的结构参数；其中，所述T为预设更新共享模型层和特有模型层的次数。
根据本公开实施例的第三方面，提供一种模型训练装置，其特征在于，应用于操作维护管理OAM实体，所述装置包括：
分组模块，用于对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组，所述无线接入网设备组包括第一数量无线接入网设备；确定模块，用于确定所述第一数量无线接入网设备对应的第一数量模型训练结构，并基于所述第一数量模型训练结构，确定第一数量特有模型层；发送模块，用于向所述第一数量无线接入网设备，发送所述第一数量特有模型层的结构参数。
一种实施方式中，所述确定模块，用于：
确定所述第一数量无线接入网设备发送的第一数量模型订阅请求，并确定所述第一数量模型订阅请求的模型训练任务特性，所述模型训练任务特性用于指示模型层数和节点数；基于所述模型训练任务特性所指示的模型层数和节点数，确定第一数量的模型训练结构。
一种实施方式中，所述确定模块，用于：
将与第一数量无线接入网设备对应的第一数量模型训练结构输出层，确定为第一数量 特有模型层。
一种实施方式中，所述确定模块，还用于：
将与第一数量无线接入网设备对应的模型训练结构输入层和隐藏层，确定为共享模型层，并获取所述多个无线接入网设备的数据，对所述数据添加与每个无线接入网设备对应的数据标识；对带有所述数据标识的所有数据进行分类处理，得到模型训练数据和模型标签值；将所述模型训练数据作为第一输入数据，输入至所述共享模型层，得到所述共享模型层输出的第一输出数据；向所述多个无线接入网设备发送所述模型标签值和所述第一输出数据。
一种实施方式中，所述装置还包括：更新模块；
所述更新模块，用于响应于接收到所述多个无线接入网设备发送的训练损失值，基于所述训练损失值更新所述共享模型层的结构参数。
一种实施方式中，所述更新模块，用于：
对所述训练损失值进行加权，得到加权损失值；确定所述共享模型层当前的模型参数和模型学习率；基于所述加权损失值、模型参数和模型学习率，确定所述共享模型层的更新参数，并基于所述更新参数对所述共享模型层的结构参数进行更新。
一种实施方式中，所述更新模块，还用于：
响应于第T次更新所述共享模型层的结构参数，确定所述共享模型层训练完成，将第T次更新后的所述共享模型层的模型结构参数，发送至所述多个无线接入网设备组中的每个无线接入网设备；其中，所述T为预设更新共享模型层和特有模型层的次数，所述共享模型层结构参数用于无线接入网设备合成无线接入网设备订阅的模型。
一种实施方式中，所述分组模块，用于：
确定每个所述模型订阅请求包括的订阅模型的类型；基于所述类型，对所述模型订阅请求进行分组，得到第一组数的模型订阅请求；对无线接入网设备进行分组，得到所述第一组数的无线接入网设备组。
一种实施方式中，所述更新模块，还用于：
响应于存在新增的无线接入网设备，且所述新增的无线接入网设备满足模型训练条件，将与新增无线接入网设备对应的特有模型层的结构参数，发送至所述新增的无线接入网设备；或，响应于存在退出的无线接入网设备，重新确定所述第一数量模型训练结构。
根据本公开实施例的第四方面，提供一种模型训练装置，其特征在于，应用于无线接入网设备，所述装置包括：
接收模块，用于接收OAM发送的特有模型层的结构参数；其中，所述特有模型层为 OAM划分第一数量模型训练结构确定的；所述第一数量模型训练结构为OAM基于无线接入网设备组包括的第一数量无线接入网设备的模型订阅请求确定的。
一种实施方式中，所述接收模块，还用于：
接收OAM发送的模型标签值和第一输出数据；将所述第一输出数据作为所述特有模型层的输入，输入至所述特有模型层，得到所述特有模型层输出的第二输出数据；基于所述模型标签值和所述第二输出数据，确定训练损失值，并向OAM发送所述训练损失值。
一种实施方式中，所述装置还包括：确定模块；
确定模块，用于基于所述模型训练数据携带的标识，在所述模型标签值中，确定与所述无线接入网设备对应的模型标签值；运算所述第二输出数据与所述训练标签值，确定训练损失值，并基于所述训练损失值对所述特有模型层的结构参数进行更新。
一种实施方式中，所述接收模块，还用于：
接收OAM发送的共享模型层的结构参数；基于所述模型共享层的结构参数和第T次更新后的特有模型层的结构参数确定订阅模型的结构参数；其中，所述T为预设更新共享模型层和特有模型层的次数。
根据本公开实施例的第五方面，提供一种模型训练装置，包括：
处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为：执行第一方面或第一方面中任意一种实施方式所述的模型训练方法，或执行第二方面或第二方面中任意一种实施方式所述的模型训练方法。
根据本公开实施例的第六方面，提供一种非临时性计算机可读存储介质，当所述存储介质中的指令由移动终端的处理器执行时，使得移动终端能够执行第一方面或第一方面中任意一种实施方式所述的模型训练方法，或使得移动终端能够执行第二方面或第二方面中任意一种实施方式所述的模型训练方法。
本公开的实施例提供的技术方案可以包括以下有益效果：
1：多个gNB-CU协同训练，OAM维护共享模型层，gNB-CU维护特有模型层，OAM必须提前告知各gNB-CU特有模型层和共享模型层的连接方式，从而gNB-CU在收到共享模型层的输出结果后，可以快速地将该输出结果输入到特有模型层中，更新模型参数。
2：OAM收集模型训练数据，需要首先向该gNB-CU组的各gNB-CU发送模型训练数据请求，之后gNB-CU会向相连接的gNB-DU发送模型训练数据请求，之后gNB-DU会向终端发送模型训练数据请求；终端在将数据发送给gNB-DU后，gNB-DU会将终端数据和gNB-DU收集的本地数据一起上报给gNB-CU，同样gNB-CU也会将gNB-DU数据和gNB-CU收集的本地数据一起上报给OAM。
3：OAM需要对模型训练数据进行数据标识，标识的信息包括该条训练数据所属的gNB-CU以及数据ID等，标识出gNB-CU信息是为了各gNB-CU能据此筛选出属于本gNB-CU的训练数据，从而利用这一部分数据来更新其特有模型层，标示出数据ID信息是为了各gNB-CU能根据ID信息检索出该条数据的标签值等信息，用于计算训练损失值。
4：在出现有gNB-CU请求加入或退出gNB-CU组的场景下，需要在gNB-CU发出请求后，OAM更新训练模型结构，同时更新gNB-CU列表，开始传输数据信息或停止传输数据信息。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，并不能限制本公开。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本公开的实施例，并与说明书一起用于解释本公开的原理。
图1是根据一示例性实施例示出的一种支持人工智能的无线网络架构示意图。
图2是根据一示例性实施例示出的一种多任务模型训练任务示意图。
图3是根据一示例性实施例示出的又一种多任务模型训练任务示意图。
图4是根据一示例性实施例示出的一种模型训练方法的系统架构示意图。
图5是根据一示例性实施例示出的又一种模型训练方法的流程图。
图6是根据一示例性实施例示出的又一种模型训练方法的流程图。
图7是根据一示例性实施例示出的一种模型训练方法的训练模型结构的流程示意图。
图8是根据一示例性实施例示出的又一种模型训练方法的流程图。
图9是根据一示例性实施例示出的又一种模型训练方法的流程图。
图10是根据一示例性实施例示出的一种训练模型结构的收集和标识模型训练数据的流程示意图。
图11是根据一示例性实施例示出的又一种模型训练方法的流程图。
图12是根据一示例性实施例示出的又一种模型训练方法的流程图。
图13是根据一示例性实施例示出的又一种模型训练方法的流程图。
图14是根据一示例性实施例示出的又一种模型训练方法的流程图。
图15是根据一示例性实施例示出的又一种模型训练方法的流程图。
图16是根据一示例性实施例示出的一种模型训练方法的新增无线接入网设备流程图。
图17是根据一示例性实施例示出的一种模型训练方法的流程图。
图18是根据一示例性实施例示出的一种模型训练方法的无线接入网设备退出流程图。
图19是根据一示例性实施例示出的又一种模型训练方法的流程图。
图20是根据一示例性实施例示出的又一种模型训练方法的流程图。
图21是根据一示例性实施例示出的又一种模型训练方法的流程图。
图22是根据一示例性实施例示出的一种模型训练方法的更新特有模型层参数流程图。
图23是根据一示例性实施例示出的又一种模型训练方法的流程图。
图24是根据一示例性实施例示出的一种模型训练方法的模型训练和模型推理示意图。
图25是根据一示例性实施例示出的一种模型训练方法和模型推理方法的流程图。
图26是根据一示例性实施例示出的一种模型训练方法的模型训练的协议和接口原理图。
图27是根据一示例性实施例示出的一种模型训练方法的模型推理的协议和接口原理图。
图28是根据一示例性实施例示出的一种模型训练方法的模型训练数据收集的协议和接口原理图。
图29是根据一示例性实施例示出的一种模型训练方法的模型推理数据收集的协议和接口原理图。
图30是根据一示例性实施例示出的一种模型训练装置框图。
图31是根据一示例性实施例示出的又一种模型训练装置框图。
图32是根据一示例性实施例示出的一种用于模型训练的装置的框图。
图33是根据一示例性实施例示出的又一种用于模型训练的装置的框图。
具体实施方式
这里将详细地对示例性实施例进行说明，其示例表示在附图中。下面的描述涉及附图时，除非另有表示，不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本公开相一致的所有实施方式。相反，它们仅是与如所附权利要求书中所详述的、本公开的一些方面相一致的装置和方法的例子。
无线网络AI架构为实现无线人工智能提供了基础，另外，根据终端具有高速移动性的场景，为保证实现模型训练和模型推理的连贯性，终端所获得的AI分析服务的连续性，以及对无线人工智能进行移动性管理，对该无线网络AI架构进行规范和优化。在3GPP会议中，提出一种支持人工智能的无线网络架构，以得到大数据使能的人工智能无线网络。
图1是根据一示例性实施例示出的一种支持人工智能的无线网络架构示意图。如图1所示，支持人工智能的无线网络架构包括数据收集/准备单元、模型训练单元、模型推理单元、行动单元、训练数据、模型性能反馈、模型部署/更新、推理数据、推理结果和性能反馈。
数据收集/准备单元：收集和AI模型训练、更新、推理相关的数据，并按照AI模型训练、更新、推理对数据内容、尺寸、格式、周期的要求，对数据进行预处理并将处理过后的数据按要求提供给模型训练单元和模型预测单元。此外，数据收集/准备单元也会根据收集的数据判断当前AI模型的有效性，并向模型训练单元提供模型性能反馈。
模型训练单元：负责AI模型的训练和更新，训练和更新所需要的输入数据由数据收集/准备单元提供，并将训练或更新好的AI模型提供给模型推理单元。
模型推理单元：基于模型训练单元提供的AI模型和数据收集/准备单元提供的输入数据，执行特定的无线网络推理任务，并将推理结果提供给行动单元和数据收集/准备单元。
行动单元：根据模型推理单元提供的推理结果，执行相应的网络行为，行动单元会收集网络侧的数据，并以性能反馈的形式提供给数据收集/准备单元。
训练数据：数据收集/准备单元对收集的数据进行预处理，将AI模型训练和更新所需要的数据提供给模型训练单元。
模型性能反馈：数据收集/准备单元根据收集的数据判断当前AI模型的有效性(例如对比预测数据和实际测量数据)，向模型训练单元提供模型性能反馈。
模型部署/更新：模型训练单元将训练和更新好的AI模型提供给模型推理单元。
推理数据：数据收集/准备单元对收集的数据进行预处理，将AI模型推理所需要的数据提供给模型推理单元。
推理结果：根据AI模型生成的推理结果，由模型推理单元提供给行动单元和数据收集/准备单元。
性能反馈：在执行相应的网络行为后，行动单元会收集网络侧的数据，并提供给数据收集/准备单元。
在相关技术中，当操作维护管理(Operation Administration and Maintenance，OAM)接收到多个训练模型的请求时，需要对每个模型训练任务进行单独训练，不同的训练任务之间相互独立。图2是根据一示例性实施例示出的一种多任务模型训练任务示意图。如图2所示，基于确定的两个模型订阅请求，首先确定两个训练数据包括训练数据1和训练数据2，基于训练数据1训练模型1，基于训练数据2训练模型2，得到任务1模型预测和模型2任务预测。在该过程中，不同的训练任务之间不会共享模型信息。
以训练单一模型、无线接入网设备为5G基站为例进一步说明。终端向5G基站分布式单元无线接入网设备(next Generation Node B Distributed Unit，gNB-DU)发起模型订阅请求，gNB-DU将该终端的模型订阅请求发送给5G基站控制单元(next Generation Node B Control Unit，gNB-CU)，gNB-CU向OAM上报终端的模型订阅请求；OAM依据终端的模型订阅 请求，选择合适的训练模型，并收集模型训练数据，进行模型训练；模型训练完成后，OAM将训练模型发送给gNB-CU，gNB-CU收集模型推理数据，进行模型推理；模型推理完成后，gNB-CU将推理模型发送给gNB-DU，gNB-DU将推理模型发送给终端；终端根据推理模型执行相应策略。
基于上述实施例训练模型的方式可知，相关技术中，存在以下技术问题：
(1)针对单个终端的模型订阅请求，进行模型训练和模型推理，最后将推理结果发送给终端，对每一个训练任务都训练一个模型，会造成较大的训练开销，没有充分考虑这些训练任务的差别和联系。
(2)OAM在进行模型训练的过程中，需要gNB-CU将数据上传到OAM，对数据的安全性造成了挑战，如果将一部分模型训练的工作转移到gNB-CU上来做，可以减少上传的数据量，降低数据安全风险。
(3)模型训练工作全部由OAM来完成，会消耗较多OAM的计算资源，如果将一部分模型训练的工作转移到gNB-CU上来做，有利于资源的均衡分配。
基于此本公开提供一种模型训练方法。图3是根据一示例性实施例示出的一种多任务模型训练任务示意图。如图3所示，基于确定的两个模型订阅请求，首先确定两个训练数据包括训练数据1和训练数据2，其中，训练数据1和训练数据2在同一个训练数据集中。基于训练数据1和训练数据2训练模型，确定任务1模型预测和任务2模型预测。
进一步的，在OAM接收到多个模型订阅请求时，将模型训练任务分配到OAM和无线接入网设备，OAM侧维护所有gNB-CU共享的共享模型层，该共享模型层用于对模型训练数据进行特征提取，gNB-CU侧维护gNB-CU独享的特有模型层，该特有模型层用于输出模型结果，将共享模型层的输出数据作为特有模型层的输入，在gNB-CU根据本地训练损失值更新特有模型层结构参数后，OAM再根据gNB-CU发送的训练损失值更新共有模型层的结构参数。协同训练的方法，不仅增加了训练模型的泛化能力，提升了用户业务体验，保障了无线网络AI分析服务的有效性，同时降低了模型训练开销，有利于提高无线网络运行效率。
进一步可以理解的是，本公开实施例的无线通信系统，是一种提供无线通信功能的网络。无线通信系统可以采用不同的通信技术，例如码分多址(code division multiple access,CDMA)、宽带码分多址(wideband code division multiple access，WCDMA)、时分多址(time division multiple access，TDMA)、频分多址(frequency division multiple access，FDMA)、正交频分多址(orthogonal frequency-division multiple access，OFDMA)、单载波频分多址(single Carrier FDMA，SC-FDMA)、载波侦听多路访问/冲突避免(Carrier Sense Multiple  Access with Collision Avoidance)。根据不同网络的容量、速率、时延等因素可以将网络分为2G(英文：generation)网络、3G网络、4G网络或者未来演进网络，如5G网络，5G网络也可称为是新无线网络(New Radio，NR)。为了方便描述，本公开有时会将无线通信网络简称为网络。
进一步的，本公开中涉及的网络设备也可以称为无线接入网设备。该无线接入网设备可以是：基站、演进型基站(evolved node B，基站)、家庭基站、无线保真(wireless fidelity，WIFI)系统中的接入点(access point，AP)、无线中继节点、无线回传节点、传输点(transmission point，TP)或者发送接收点(transmission and reception point，TRP)等，还可以为NR系统中的gNB，或者，还可以是构成基站的组件或一部分设备等。当为车联网(V2X)通信系统时，网络设备还可以是车载设备。应理解，本公开的实施例中，对网络设备所采用的具体技术和具体设备形态不做限定。
进一步的，本公开中涉及的终端，也可以称为终端设备、用户设备(User Equipment，UE)、移动台(Mobile Station，MS)、移动终端(Mobile Terminal，MT)等，是一种向用户提供语音和/或数据连通性的设备，例如，终端可以是具有无线连接功能的手持式设备、车载设备等。目前，一些终端的举例为：智能手机(Mobile Phone)、口袋计算机(Pocket Personal Computer，PPC)、掌上电脑、个人数字助理(Personal Digital Assistant，PDA)、笔记本电脑、平板电脑、可穿戴设备、或者车载设备等。此外，当为车联网(V2X)通信系统时，终端设备还可以是车载设备。应理解，本公开实施例对终端所采用的具体技术和具体设备形态不做限定。
图4是根据一示例性实施例示出的一种模型训练方法的系统架构示意图。如图4所示，该系统包括终端，无线接入网设备(其中，无线接入网设备包括gNB-DU和gNB-CU)以及OAM。终端通过无线信道接入gNB-DU，多个gNB-DU通过F1接口接入gNB-CU，gNB-CU之间通过Xn接口连接。OAM主要承担支持AI的无线网络架构中模型训练功能单元和数据收集/准备功能单元的工作，主要负责共享模型训练和数据收集的工作。gNB-CU主要承担支持AI的无线网络架构中模型训练功能单元、数据收集/准备功能单元和模型推理功能单元的工作，主要负责特有模型训练、模型推理和数据收集的工作。gNB-DU主要承担支持AI的无线网络架构中数据收集/准备功能单元的工作，主要负责数据收集的工作。终端主要承担支持AI的无线网络架构中行动执行功能单元的工作，主要负责策略执行和性能反馈的工作。
基于上述模型训练方法的系统架构示意图，本公开提供一种模型训练方法，下述实施例将结合附图对模型训练方法进行说明。
图5是根据一示例性实施例示出的一种模型训练方法的流程图。如图5所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S11中，对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组。
其中，无线接入网设备组包括第一数量无线接入网设备。
在本公开实施例中，OAM接收多个无线接入网设备发送的多个模型订阅请求，确定每个模型订阅请求中包括的终端的标识，模型请求类型、接入位置等信息。其中，终端标识为全球唯一临时用户设备标识(Globally Unique Temporary UE Identity，GUTI)。模型请求类型以分析ID来表示，如负载预测分析服务。接入位置信息主要包含终端当前接入的gNB-CU和gNB-DU信息。
OAM根据以上信息，分析模型请求中请求的模型的相似度，根据相似度对不同无线接入网设备发送的模型订阅请求进行分组。例如gNB-CU(n1)到gNB-CU(n2)的分析订阅请求信息显示终端请求用于负载预测的模型，gNB-CU(n3)到gNB-CU(n4)的分析订阅请求信息显示终端请求用于网络决策的模型，根据不同训练任务之间的相似度，可以将gNB-CU(n1)到gNB-CU(n2)划分为一组进行模型训练，将gNB-CU(n3)到gNB-CU(n4)划分为另外一组进行模型训练。
在步骤S12中，确定第一数量无线接入网设备对应的第一数量模型训练结构，并基于第一数量模型训练结构，确定第一数量特有模型层。
在本公开实施例中，OAM为每个无线接入网设备组确定模型训练结构，即，为每个无线接入网设备组中第一数量无线接入网设备确定对应的第一数量模型训练结构。将第一数量模型训练结构划分为共享模型层和特有模型层。确定每个模型训练结构对应的特有模型层。
在步骤S13中，向第一数量无线接入网设备，发送第一数量特有模型层的结构参数。
在本公开实施例中，将确定的每个模型训练结构对应的特有模型层，发送至相应的无线接入网设备。以无线接入设备为gNB-CU为例，OAM按照各gNB-CU到各连接方式的映射表，向每一个gNB-CU发送其特有模型层的结构参数信息；各gNB-CU接收模型信息，将特有模型层作为本地模型进行模型训练和更新。
通过本公开实施例提供的模型训练方法，可以实现将一部分模型训练的工作转移到无线接入网设备，可以减少上传的数据量，有利于资源的均衡分配，降低数据安全风险。
图6是根据一示例性实施例示出的一种模型训练方法的流程图。如图6所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S21中，确定第一数量无线接入网设备发送的第一数量模型订阅请求，并确定第一数量模型订阅请求的模型训练任务特性。
其中，模型训练任务特性用于指示模型层数和节点数。
在本公开实施例中，模型订阅请求的模型训练任务不同，对应请求训练的模型的层数和节点数也不相同，可以基于模型训练任务确定每个模型订阅请求的模型训练任务特性。包括，将输入层节点数设置为M，代表一次输入模型中的训练数据数量。输出层节点数设置为N，N取决于gNB-CU的个数和训练任务特性，例如预测任务(回归任务)中每个gNB-CU对应一个节点，决策任务中每个gNB-CU对应多个节点。隐藏层数量设置为S，每个隐藏层的节点数设置为L，隐藏层的数量需要考虑模型大小和模型泛化能力等因素。从耳确定模型训练任务特性。
在步骤S22中，基于模型训练任务特性所指示的模型层数和节点数，确定第一数量的模型训练结构。
在本公开实施例中，模型训练结构包括模型层与层之间的连接方式，对应层之间的连接方式，其中隐藏层与输入层之间可以是全连接方式，激活函数可以使用ReLU函数。隐藏层与隐藏层之间可以是全连接方式，激活函数可以使用ReLU函数。隐藏层与输出层之间可以是部分连接方式，激活函数可以使用Softmax函数或Sigmoid函数。
其中需要说明的是，对应所使用的损失函数的确定过程，可以参考预测任务(回归任务)采用均方误差(MSE)损失函数、平均绝对误差(MAE)损失函数、Huber损失函数等，决策任务(分类任务)采用交叉熵损失函数、Hinge损失函数、对数损失函数等。
对应网络模型的超参数的确定过程，可以参考学习轮次设置为T次，学习轮次的设置需要衡量模型训练速度和训练成本以及模型训练精度的影响，学习率设置为α和β。权重初始化的方法选择随机权重初始化。
图7是根据一示例性实施例示出的一种模型训练方法的训练模型结构的流程示意图。如图7所示，包括以下步骤：
步骤S211，首先OAM根据gNB-CU组的gNB-CU数量以及训练任务特性等信息，确定训练模型的模型结构。
步骤S212，OAM将训练模型划分为共享模型层和特有模型层，其中共享模型层是所有gNB-CU共同使用，特有模型层是每个gNB-CU单独使用。
步骤S213，OAM确定共享模型层和特有模型层的连接方式，并初始化模型参数。
步骤S214，OAM将每个gNB-CU的特有模型层的结构和参数发送给对应gNB-CU。
图8是根据一示例性实施例示出的一种模型训练方法的流程图。如图8所示，模型训 练方法用于OAM中，包括以下步骤。
在步骤S31中，将与第一数量无线接入网设备对应的第一数量模型训练结构输出层，确定为第一数量特有模型层。
在本公开实施例中，确定第一数据无线接入网设备中每个无线接入网设备对应的模型训练结构的输出层，将该输出层确定为特有模型层，得到第一数量的特有模型层。其中，每个特有模型层为对应无线接入网设备单独使用，用于输出最终分类或回归的结果。
图9是根据一示例性实施例示出的一种模型训练方法的流程图。如图9所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S41中，将与第一数量无线接入网设备对应的模型训练结构输入层和隐藏层，确定为共享模型层，并获取多个无线接入网设备的数据，对数据添加与每个无线接入网设备对应的数据标识。
在本公开实施例中，确定第一数据无线接入网设备中每个无线接入网设备对应的模型训练结构的输入层和隐藏层，将该输入层和隐藏层确定为共享模型层。其中共享模型层是所有gNB-CU共同使用，用于提取输入数据的特征信息。
在本公开实施例中，OAM还可以向无线接入网设备请求获取模型训练数据，进行数据处理，并对每条训练数据进行标识，标识出数据所属于的无线接入网设备以及数据ID等信息。进一步地，OAM向所管理的各无线接入网设备发送模型训练数据请求。无线接入网设备接收模型训练数据请求后，向所连接的各无线接入网设备发送模型训练数据请求。各无线接入网设备向接入该无线接入网设备的终端发送模型训练数据请求，终端接收模型训练数据请求后，收集终端数据发送给无线接入网设备。无线接入网设备汇总终端训练数据，并收集本无线接入网设备的数据，发送给与其连接的无线接入网设备。无线接入网设备汇总无线接入网设备数据，并收集本无线接入网设备的数据，发送给OAM。
其中，无线接入网设备信息可以用0-1编码方式来标识，假设无线接入网设备数量为N，则使用N比特来记录无线接入网设备信息，每一比特如果为0代表该数据不属于对应gNB-CU，如果为1代表该数据属于对应无线接入网设备；数据ID信息需要与无线接入网设备处的数据ID信息保持一致。
在步骤S42中，对带有数据标识的所有数据进行分类处理，得到模型训练数据和模型标签值。
在本公开实施例中，OAM对模型训练数据进行数据处理，如数据去噪、归一化等，得到用于模型训练的数据，包括共享模型层的模型训练数据和特有模型层的模型标签值。
通过本公开实施例提供的模型训练方法，可以扩充训练任务对应的数据量，实现部分 数据共享。并且共享数据在一定程度上弱化了网络能力，降低了过拟合的风险。
图10是根据一示例性实施例示出的一种训练模型结构的收集和标识模型训练数据的流程示意图。如图10所示，以无线接入网设备为gNB-CU为例，包括以下步骤：
步骤S421，OAM向各gNB-CU发送收集数据请求。
步骤S422，各gNB-CU收集终端数据，发送给OAM。
步骤S423，OAM汇总各gNB-CU发送的数据，形成模型训练数据。
步骤S424，OAM对模型训练数据进行数据处理，如数据去噪、归一化等。
步骤S425，OAM对模型训练数据进行数据标识，标识每条数据记录所属于的gNB-CU信息，以及对应的ID信息等。
在步骤S43中，将模型训练数据作为第一输入数据，输入至共享模型层，得到共享模型层输出的第一输出数据。
在本公开实施例中，OAM将模型训练数据作为共享模型层的输入，OAM以串行的方式，将模型训练数据输入至共享模型层，共享模型层的最后一层有L个节点，因此每条模型训练数据i对应一组输出结果
Figure PCTCN2021098008-appb-000001
得到共享模型层的所有输出结果，即共享模型层输出的第一输出数据，本公开为便于描述将共享模型层的输出结果成为第一输出数据。
在步骤S44中，向多个无线接入网设备发送模型标签值和第一输出数据。
在本公开实施例中，在得到模型训练数据的全部输出结果(即，第一输出数据)后，OAM将第一输出数据以及模型训练数据的标识信息模型标签值发送给各无线接入网设备。
通过本公开提供的模型训练数据可以使得多模型训练任务协同进行，会互相增加噪声，从而提高模型的泛化能力。
图11是根据一示例性实施例示出的一种模型训练方法的流程图。如图11所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S51中，响应于接收到多个无线接入网设备发送的训练损失值，基于训练损失值更新共享模型层的结构参数。
在本公开实施例中，OAM接收到多个无线接入网设备发送的训练损失值后，依据各损失值对共享模型层结构参数进行更新。
图12是根据一示例性实施例示出的一种模型训练方法的流程图。如图12所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S61中，对训练损失值进行加权，得到加权损失值。
在本公开实施例中，若当前为多个无线接入网设备第t次发送的训练损失值，则OAM基于各gNB-CU的数据量以及学习效果等因素，对训练损失值进行加权。其中，对训练损 失值进行加权可参考如下公式：
Figure PCTCN2021098008-appb-000002
其中，
Figure PCTCN2021098008-appb-000003
为训练损失值，K为gNB-CU数量，w k为第k个gNB-CU训练损失值的权重。
其中，w k的计算包括两方面，一方面为每个gNB-CU的训练数据量占总数据量的比重，另一方面为学习效果的影响，例如训练模型的精确度、学习任务的难易程度等。
在步骤S62中，确定共享模型层当前的模型参数和模型学习率。
在步骤S63中，基于加权损失值、模型参数和模型学习率，确定共享模型层的更新参数，并基于更新参数对共享模型层的结构参数进行更新。
在本公开实施例中，OAM确定使用加权后的训练损失值、模型更新方法以及选择的结构参数，对共享模型层结构参数进行更新。例如，采用SGD算法对共享模型层参数进行更新，可参见如下公式：
Figure PCTCN2021098008-appb-000004
其中，b t表示第t轮待更新的共享模型层结构参数，loss t表示第t轮计算得到的加权训练损失值，β t表示第t轮的学习率。
图13是根据一示例性实施例示出的一种模型训练方法的流程图。如图13所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S71中，响应于第T次更新共享模型层的结构参数，确定共享模型层训练完成，将第T次更新后的共享模型层的模型结构参数，发送至多个无线接入网设备组中的每个无线接入网设备。
其中，T为预设更新共享模型层和特有模型层的次数，共享模型层结构参数用于无线接入网设备合成无线接入网设备订阅的模型。
在本公开实施例中，OAM在第T次完成共享模型层的结构参数更新之后，将第T次更新共享模型层的结构参数，发送至各个无线接入网设备。无线接入网设备在接收到模型信息后，由于无线接入网设备保存有共享模型层与特有模型层的连接方式，因此无线接入网设备可以将两个模型按照特定的连接方式拼接起来，整合成一个完整的模型，用于模型推理。
图14是根据一示例性实施例示出的一种模型训练方法的流程图。如图14所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S81中，确定每个模型订阅请求包括的订阅模型的类型。
在本公开实施例中，OAM确定各个无线接入网设备发送的模型订阅请求中包括的请 求模型训练任务的类型，例如是负载预测模型训练任务，网络决策训练任务等。
在步骤S82中，基于类型，对模型订阅请求进行分组，得到第一组数的模型订阅请求。
在本公开实施例中，OAM基于模型训练任务的类型，基于模型训练任务类型的相似度，对接收的模型订阅请求进行分组，确定不同的多组模型订阅请求，进一步得到第一组数的模型订阅请求。
在步骤S83中，对无线接入网设备进行分组，得到第一组数的无线接入网设备组。
在本公开实施例中，OAM根据第一组数的模型订阅请求对相应的无线接入网设备进行分组，得到第一组数的无线接入网设备组。
通过本公开对无线接入网设备进行分组，可以提高训练效率。并且OAM采用协同训练的训练方法，在协同训练的方法下，参与训练的各无线接入网设备的训练任务越相似，训练效果越好，因此将无线接入网设备分组进行训练。
图15是根据一示例性实施例示出的一种模型训练方法的流程图。如图15所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S91中，响应于存在新增的无线接入网设备，且新增的无线接入网设备满足模型训练条件，将与新增无线接入网设备对应的特有模型层的结构参数，发送至新增的无线接入网设备。
在本公开实施例中，当新无线接入网设备请求加入无线接入网设备组时，新无线接入网设备首先向OAM发送加入当前gNB-CU组的请求。OAM收到请求后，判断新无线接入网设备是否满足加入当前无线接入网设备组的条件，若满足条件，则可以加入当前无线接入网设备组参与模型训练，若不满足条件，则不可以加入当前无线接入网设备组。若满足条件，OAM更新无线接入网设备组的信息，将新无线接入网设备加入到参与模型训练的无线接入网设备列表中。OAM根据新无线接入网设备的训练任务特性等信息，更新训练模型结构，并将新无线接入网设备的特有模型层结构和参数发送给新无线接入网设备。OAM向新无线接入网设备发送共享模型层输出结果，新无线接入网设备更新特有模型层结构参数，进行模型训练。
一种实施例中，以无线接入网设备为gNB-CU为例进行说明。图16是根据一示例性实施例示出的一种模型训练方法的新增无线接入网设备流程图。如图16所示，包括以下步骤：
步骤S911，新gNB-CU首先向OAM发送加入当前gNB-CU组的请求。
步骤S912，OAM收到请求后，首先判断新gNB-CU是否满足加入当前gNB-CU组的条件，若满足条件，则可以加入当前gNB-CU组参与模型训练，若不满足条件，则不可以 加入当前gNB-CU组。
步骤S913，若满足条件，OAM更新gNB-CU组的信息，将新gNB-CU加入到参与模型训练的gNB-CU列表中。
步骤S914，OAM根据新gNB-CU的训练任务特性等信息，更新训练模型结构，并将新gNB-CU的特有模型层结构和参数发送给新gNB-CU。
步骤S915，OAM向新gNB-CU发送共享模型层输出结果，新gNB-CU更新特有模型层参数，进行模型训练。
在本公开一些实施例中，在当前gNB-CU组进行训练的过程中，存在新的终端向gNB-CU发送分析订阅请求，进而有新的gNB-CU向OAM发送模型订阅请求。OAM在收到新gNB-CU发送的请求后，首先判断该gNB-CU是否有加入当前gNB-CU组的条件。一种实施例中，判断新gNB-CU是否有加入当前gNB-CU组的条件的方法是，比较新gNB-CU的终端模型订阅请求信息中分析请求类型与本gNB-CU组的终端模型订阅请求信息中分析请求类型的相似度，如果相似度较高，则新gNB-CU满足条件，可以加入当前gNB-CU组参与模型训练，如果相似度较低，则新gNB-CU不满足条件，不可以加入当前gNB-CU组参与模型训练。例如新gNB-CU请求训练模型执行预测任务，而本gNB-CU组训练模型用于决策任务，则两者相似度较低，该gNB-CU不可以加入当前gNB-CU组。如果新gNB-CU满足条件，则OAM将其添加到参加模型训练的列表中，并更新gNB-CU组的信息，开始向其发送数据信息。OAM在现有训练模型的基础上，在不更改共享模型层的条件下，修改特有模型层的结构，包括添加分支、增加输出层节点数、更改与共享模型层的连接方式等，对训练模型结构进行更新，并将新增加的特有模型层结构和参数发送给新gNB-CU，作为新gNB-CU的特有模型层。
图17是根据一示例性实施例示出的一种模型训练方法的流程图。如图17所示，模型训练方法用于OAM中，包括以下步骤。
在步骤S101中，响应于存在退出的无线接入网设备，重新确定第一数量模型训练结构。
在本公开实施例中，当无线接入网设备请求退出无线接入网设备组时，该无线接入网设备首先向OAM发送退出当前无线接入网设备组的请求。OAM收到退出请求后，在参与模型训练的无线接入网设备列表中删除该无线接入网设备的相关信息，不再向该无线接入网设备发送数据；OAM删除该无线接入网设备的特有模型层，更新训练模型结构。该无线接入网设备不再参加该无线接入网设备组的模型训练过程，若该无线接入网设备还没有完成当前轮次模型训练，则无线接入网设备继续完成当前轮次模型训练，但不再进行参数 上传。
一种实施例中，以无线接入网设备为gNB-CU为例进行说明。图18是根据一示例性实施例示出的一种模型训练方法的无线接入网设备退出流程图。如图18所示，包括以下步骤：
步骤S1011，gNB-CU向OAM发送退出当前gNB-CU组的请求。
步骤S1012，OAM收到退出请求后，在参与模型训练的gNB-CU列表中删除该gNB-CU的相关信息，不再向该gNB-CU发送数据。
步骤S1013，OAM删除该gNB-CU的特有模型层，更新训练模型结构。
步骤S1014，该gNB-CU不再参加该gNB-CU组的模型训练过程。
在本公开一些实施例中，在当前gNB-CU组进行训练的过程中，存在终端向gNB-CU取消分析订阅请求，进而有gNB-CU向OAM发送取消模型订阅的请求，请求退出当前gNB-CU组。OAM收到请求后，将该gNB-CU从参加模型训练的列表中删除，并更新gNB-CU组的信息，不再向其发送数据信息。OAM在现有训练模型的基础上，在不更改共享模型层的条件下，修改特有模型层的结构，包括删除分支、减少输出层节点数、更改与共享模型层的连接方式等，对训练模型结构进行更新，删除该gNB-CU的特有模型层结构。该gNB-CU不再参加该gNB-CU组的模型训练过程。若该gNB-CU还没有完成当前轮次模型训练，则gNB-CU继续完成当前轮次模型训练，但不再进行参数上传。
基于相同的构思，本公开实施例还提供一种模型训练方法。
图19是根据一示例性实施例示出的一种模型训练方法的流程图。如图19所示，模型训练方法用于无线接入网设备中，包括以下步骤。
在步骤S111中，接收OAM发送的特有模型层的结构参数。
在本公开实施例中，无线接入网设备接收OAM发送的特有模型层。特有模型层为OAM划分第一数量模型训练结构确定的；第一数量模型训练结构为OAM基于无线接入网设备组包括的第一数量无线接入网设备的模型订阅请求确定的。
其中，以无线接入设备为gNB-CU为例，OAM按照各gNB-CU到各连接方式的映射表，向每一个gNB-CU发送其特有模型层的结构参数信息；各gNB-CU接收模型信息，将特有模型层作为本地模型进行模型训练和更新。
通过本公开实施例提供的模型训练方法，可以实现将一部分模型训练的工作转移到无线接入网设备，可以减少上传的数据量，有利于资源的均衡分配，降低数据安全风险。
图20是根据一示例性实施例示出的一种模型训练方法的流程图。如图20所示，模型训练方法用于无线接入网设备中，包括以下步骤。
在步骤S121中，接收OAM发送的模型标签值和第一输出数据。
在本公开实施例中，无线接入网设备接收OAM发送的第一输出数据和模型标签值，各无线接入网设备根据模型标签值的无线接入网设备标识信息，确定属于该无线接入网设备的模型标签值，并筛选出这些模型标签值的共享模型层输出的第一输出数据。
在步骤S122中，将第一输出数据作为特有模型层的输入，输入至特有模型层，得到特有模型层输出的第二输出数据。
在本公开实施例中，无线接入网设备将筛选与自身对应的共享模型层输出的第一输出数据，输入至该无线接入网设备接收的特有模型层中，得到特有模型层的第二输出数据。
其中，各无线接入网设备保存有OAM发送的特有模型层的结构和参数信息，其中包括特有模型层和共享模型层的连接方式，各无线接入网设备根据特有模型层和共享模型层的连接方式，可以将共享模型层的输出结果输入到特有模型层中。
一种实施例中，无线接入网设备将经过判断属于本无线接入网设备的共享模型层的输出结果，串行地输入到特有模型层中，得到特有模型层输出结果，输出层节点数为N，因此每条模型训练数据i对应一组输出结果
Figure PCTCN2021098008-appb-000005
在步骤S123中，基于模型标签值和第二输出数据，确定训练损失值，并向OAM发送训练损失值。
在本公开实施例中，无线接入网设备确定与自身对应的模型标签值，根据模型标签值和第二输出数据，确定训练损失值。其中训练损失值的确定可以参考以下公式：
Figure PCTCN2021098008-appb-000006
其中，loss为训练损失值，I为属于本gNB-CU的模型训练数据量，y i为数据i经过特有模型层的输出结果，
Figure PCTCN2021098008-appb-000007
为数据i的标签值。
在本公开一些实施例中，还可以根据不同任务选择适当的损失函数。
图21是根据一示例性实施例示出的一种模型训练方法的流程图。如图21所示，模型训练方法用于无线接入网设备中，包括以下步骤。
在步骤S131中，基于模型训练数据携带的标识，在模型标签值中，确定与无线接入网设备对应的模型标签值。
在本公开实施例中，各无线接入网设备根据经过判断属于自身的模型训练数据的ID信息，在该无线接入网设备的数据库中进行查找，获得每一条模型训练数据的标签值等信息。无线接入网设备在接收模型训练数据的标识信息后，通过解析其携带的N比特的无线 接入网设备信息，判断该训练数据是否属于本无线接入网设备，进而确定所有属于本无线接入网设备的训练数据，获得其数据ID信息以及经过共享模型层的输出结果。
在步骤S132中，运算第二输出数据与所述训练标签值，确定训练损失值，并基于训练损失值对特有模型层的结构参数进行更新。
在本公开实施例中，各gNB-CUk使用各自得到的训练损失值、模型更新方法以及选择的结构参数，例如随机梯度下降算法(Stochastic Gradient Descent，SGD)、优化算法Adam等，对特有模型层参数进行更新，例如采用SGD算法对特有模型层参数进行更新，可参见如下公式：
Figure PCTCN2021098008-appb-000008
其中，
Figure PCTCN2021098008-appb-000009
表示第t轮待更新的特有模型层参数，
Figure PCTCN2021098008-appb-000010
表示第t轮更新后的特有模型层参数，
Figure PCTCN2021098008-appb-000011
表示第t轮计算得到的训练损失值的梯度，
Figure PCTCN2021098008-appb-000012
表示第t轮的学习率。
一种实施例中，以无线接入网设备为gNB-CU为例进行说明。图22是根据一示例性实施例示出的一种模型训练方法的更新特有模型层参数流程图。如图22所示，包括以下步骤：
步骤S1321，各gNB-CU根据训练数据的gNB-CU标识信息，确定哪些训练数据属于本gNB-CU，筛选出这些数据的共享模型层输出结果。
步骤S1322，各gNB-CU根据训练数据的标识信息，获得训练数据的标签值等信息。
步骤S1323，各gNB-CU以共享模型层的输出结果作为特有模型层的输入，得到特有模型层输出结果。
步骤S1324，各gNB-CU根据特有模型层的输出结果以及标签值信息，计算训练损失值。
步骤S1325，各gNB-CU根据训练损失值，对其特有模型层参数进行更新。
图23是根据一示例性实施例示出的一种模型训练方法的流程图。如图23所示，模型训练方法用于无线接入网设备中，包括以下步骤。
在步骤S141中，接收OAM发送的共享模型层的结构参数。
在步骤S142中，基于模型共享层的结构参数和第T次更新后的特有模型层的结构参数确定订阅模型的结构参数。
其中，T为预设更新共享模型层和特有模型层的次数。
在本公开实施例中，OAM将共享模型层的模型参数发送给无线接入网设备组的每一个无线接入网设备，无线接入网设备在接收到共享模型层的结构参数后，无线接入网设备基于保存的共享模型层与特有模型层的连接方式，将两个模型的结构参数按照特定的连接 方式拼接起来，整合成一个完整的模型，得到该模型的结构参数，可以用于模型推理。
在本公开一些实施例中，无线接入网设备确定完整的模型之后，可以将该完整的模型用于模型推理，其模型推理过程包括：
(1)无线接入网设备收集模型推理数据，基于推理模型进行模型推理，将推理结果发送给终端，同时将模型推理结果反馈给OAM。
进一步地，无线接入网设备向所连接的各无线接入网设备发送模型训练数据请求。各无线接入网设备向接入该无线接入网设备的所有终端发送模型训练数据请求。终端接收模型训练数据请求后，收集终端数据发送给无线接入网设备。无线接入网设备汇总终端数据，并收集本无线接入网设备的数据，发送给与其连接的无线接入网设备。无线接入网设备汇总无线接入网设备数据，并收集本无线接入网设备的数据，形成模型推理数据。
(2)无线接入网设备在收集完成模型推理数据后，基于整合后的推理模型进行模型推理，得到推理结果发送给终端。
进一步地，无线接入网设备将推理结果发送给终端接入的无线接入网设备。无线接入网设备将推理结果发送给终端。无线接入网设备完成模型推理后，将推理结果反馈给OAM。其中，无线接入网设备需要反馈的推理结果为模型推理的准确度等信息。
(3)终端根据模型推理结果，执行网络优化策略，收集网络性能数据并反馈给OAM，用于模型训练。
其中，终端根据模型推理结果(例如预测结果、决策结果等)，执行相应的网络优化策略(例如小区切换、小区激活等)；同时终端会收集网络侧的性能数据(例如测量结果，小区切换成功或失败的相关数据等)，并反馈给无线接入网设备，无线接入网设备将性能数据反馈给OAM，用于模型训练。
在本公开一些实施例中，协同模型训练和模型推理的过程可参见图24。图24是根据一示例性实施例示出的一种模型训练方法的模型训练和模型推理示意图。如图24所示，包括共享模型层和特有模型层的结构，其中共享模型层包括输入层和隐藏层，特有模型层包括输出层，共享模型层在OAM上进行训练，特有模型层在无线接入网设备(例如gNB-CU)上进行训练，每一个gNB-CU只保留特有模型层的一个分支。
OAM得到共享模型层的输出结果，并将其发送给每一个gNB-CU；每一个gNB-CU再根据共享模型层的输出结果，获得其特有模型层的输出结果，并计算训练损失值，更新特有模型层的模型参数；每一个gNB-CU将训练损失值发送给OAM，OAM更新共享模型层的模型参数，继续进行训练。OAM将训练完成的共享模型层的模型信息发送给每一个gNB-CU；每一个gNB-CU将共享模型层与本地特有模型层整合，形成完整的推理模型； 每一个gNB-CU使用推理模型进行模型推理。
在本公开一些实施例中，将结合OAM，无线接入网设备(例如，gNB-CU)终端之间的交互，对模型训练以及模型推理过程进行说明。图25是根据一示例性实施例示出的一种模型训练方法和模型推理方法的流程图。如图25所示，包括以下步骤：
步骤S151，终端向所属gNB-CU发起分析订阅请求，各gNB-CU根据各自终端的请求获取本地训练的任务特性等信息。
步骤S152，各gNB-CU向OAM发送模型订阅请求。
步骤S153，OAM汇总各gNB-CU的模型订阅请求，根据训练任务的相似度对gNB-CU进行分组，得到不同的gNB-CU组。
步骤S154，对于每个gNB-CU组，OAM确定合适的训练模型结构，将训练模型划分为共享模型层和特有模型层，初始化模型参数，并将各gNB-CU的特有模型层结构参数发送给相应gNB-CU。
步骤S155，OAM收集模型训练数据，进行数据处理，并对每条训练数据进行标识，标识出数据所属于的gNB-CU以及数据ID等信息。
步骤S156，OAM将模型训练数据作为输入，得到共享模型层的输出结果，将输出结果以及对应训练数据的标识信息发送给各gNB-CU。
步骤S157，各gNB-CU根据训练数据的标识信息进行筛选并获得训练数据的模型标签值，以共享模型层的输出结果作为特有模型层的输入，得到输出结果并计算训练损失值，对特有模型层参数进行更新。
步骤S158，各gNB-CU向OAM发送训练损失值，OAM依据各损失值对共享模型层结构参数进行更新。
步骤S159，当gNB-CU需要加入或退出gNB-CU组时，gNB-CU向OAM发出加入或退出的请求，并完成加入或退出gNB-CU组的流程。
步骤S160，模型训练完成后，OAM将共享模型层结构参数发送给各gNB-CU，gNB-CU接收模型结构参数并进行整合，形成完整的推理模型。
步骤S161，gNB-CU收集模型推理数据，基于推理模型进行模型推理，将推理结果发送给终端，同时将模型推理结果反馈给OAM。
步骤S162，终端根据模型推理结果，执行网络优化策略，收集网络性能数据，反馈给gNB-CU和OAM，用于模型训练。
需要说明的是，该实施例对于OAM，无线接入网设备(例如，gNB-CU)终端之间交互过程的说明可以参见上述实施例，在此不再赘述。
图26是根据一示例性实施例示出的一种模型训练方法的模型训练的协议和接口原理图。如图26所示，主要涉及本公开实施例提供的终端、终端接入的无线接入网设备(例如gNB-DU)、终端接入的无线接入网设备(例如gNB-CU)以及OAM。具体如下：
1a.终端将分析订阅请求信令发送给gNB-DU。1b.gNB-DU接收终端发送的分析订阅请求信令，并将分析订阅请求信令发送给gNB-CU。2.gNB-CU接收gNB-DU发送的分析订阅请求信令，并形成模型订阅请求信令。3.gNB-CU将模型订阅请求信令发送给OAM。4.OAM接收模型订阅请求信令，获取信令中包含的信息，对gNB-CU进行分组。5.OAM将分组信息发送给本gNB-CU组的各gNB-CU。6.OAM确定训练模型结构，将模型划分为共享模型层和特有模型层。7.OAM将特有模型层的模型信息发送给本gNB-CU组的各gNB-CU。8.各gNB-CU接收特有模型层的模型信息，用于模型训练。9.OAM收集模型训练数据，完成数据收集后对数据进行处理和标识。10.OAM将模型训练数据作为共享模型层输入，得到输出结果。11.OAM向各gNB-CU发送共享模型层输出结果和训练数据的标识信息。12.各gNB-CU根据训练数据的标识信息，筛选出属于本gNB-CU的训练数据对应的输出结果，并获得这些训练数据的标签值。13.各gNB-CU以筛选后的共享模型层的输出结果，作为特有模型层的输入，得到特有模型层的输出结果，并计算训练损失值，对特有模型层参数进行更新。14.各gNB-CU向OAM发送计算的训练损失值。15.OAM接收汇总各训练损失值，对共享模型层参数进行更新。
图27是根据一示例性实施例示出的一种模型训练方法的模型推理的协议和接口原理图。如图27所示，主要涉及本发明实施例提供的终端、终端接入的无线接入网设备(例如gNB-DU)、终端接入的无线接入网设备(例如gNB-CU)以及OAM。具体如下：
1.各gNB-CU将模型请求信令发送给OAM。2.OAM接收模型请求信令，准备共享模型层的模型信息。3.OAM将共享模型层的模型信息发送给各gNB-CU。4.各gNB-CU接收共享模型层信息，与gNB-CU本地的特有模型层进行整合，形成完整的推理模型。5.各gNB-CU收集模型推理数据，基于推理模型进行模型推理，得到推理结果。6a.各gNB-CU将模型推理结果发送给相连接的gNB-DU。6b.各gNB-DU将模型推理结果发送给相连接的终端。6c.各gNB-CU将模型推理反馈结果发送给OAM。7.终端根据模型推理结果，执行网络优化策略，并收集网络性能数据。8a.终端将网络性能数据反馈给相连接的gNB-DU。8b.gNB-DU将网络性能数据反馈给相连接的gNB-CU。8c.gNB-CU将网络性能数据反馈给OAM。9.gNB-CU和OAM接收网络性能反馈数据，用于模型训练。
图28是根据一示例性实施例示出的一种模型训练方法的模型训练数据收集的协议和接口原理图。如图28所示，主要涉及本发明实施例提供的终端、终端接入的无线接入网 设备(例如gNB-DU)、终端接入的无线接入网设备(例如gNB-CU)以及OAM。
1a.OAM将模型训练数据请求信令发送给各gNB-CU。1b.各gNB-CU将模型训练数据请求信令发送给相连接的gNB-DU。1c.各gNB-DU将模型训练数据请求信令发送给相连接的终端。2.终端接收模型训练数据请求，准备终端训练数据。3.终端将训练数据发送给相连接的gNB-DU。4.各gNB-DU接收终端训练数据，并收集本gNB-DU的数据，形成gNB-DU训练数据。5.各gNB-DU将训练数据发送给相连接的gNB-CU。6.各gNB-CU接收gNB-DU训练数据，并收集本gNB-CU的数据，形成gNB-CU训练数据。7.各gNB-CU将训练数据发送给OAM。8.OAM接收gNB-CU训练数据，并收集OAM本地数据，形成模型训练数据。
图29是根据一示例性实施例示出的一种模型训练方法的模型推理数据收集的协议和接口原理图。如图29所示，主要涉及本发明实施例提供的终端、终端接入的无线接入网设备(例如gNB-DU)、终端接入的无线接入网设备(例如gNB-CU)以及OAM。
1a.当前gNB-CU将模型推理数据请求信令发送给相连接的gNB-DU。1b.各gNB-DU将模型推理数据请求信令发送给相连接的终端。2.终端接收模型推理数据请求，准备终端推理数据。3.终端将推理数据发送给相连接的gNB-DU。4.各gNB-DU接收终端推理数据，并收集本gNB-DU的数据，形成gNB-DU推理数据。5.各gNB-DU将推理数据发送给相连接的gNB-CU。6.当前gNB-CU接收gNB-DU推理数据，并收集本gNB-CU的数据，形成模型推理数据。
基于相同的构思，本公开实施例还提供一种模型训练装置。
可以理解的是，本公开实施例提供的模型训练装置为了实现上述功能，其包含了执行各个功能相应的硬件结构和/或软件模块。结合本公开实施例中所公开的各示例的单元及算法步骤，本公开实施例能够以硬件或硬件和计算机软件的结合形式来实现。某个功能究竟以硬件还是计算机软件驱动硬件的方式来执行，取决于技术方案的特定应用和设计约束条件。本领域技术人员可以对每个特定的应用来使用不同的方法来实现所描述的功能，但是这种实现不应认为超出本公开实施例的技术方案的范围。
图30是根据一示例性实施例示出的一种模型训练装置框图。参照图30，该装置100应用于操作维护管理OAM实体，包括分组模块101，确定模块102和发送模块103。
分组模块101，用于对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组，无线接入网设备组包括第一数量无线接入网设备。确定模块102，用于确定第一数量无线接入网设备对应的第一数量模型训练结构，并基于第一数量模型训练结构，确定第一数量特有模型层。发送模块103，用于向第一数量无线接入网设备，发 送第一数量特有模型层的结构参数。
在本公开实施例中，确定模块102，用于确定第一数量无线接入网设备发送的第一数量模型订阅请求，并确定第一数量模型订阅请求的模型训练任务特性，模型训练任务特性用于指示模型层数和节点数。基于模型训练任务特性所指示的模型层数和节点数，确定第一数量的模型训练结构。
在本公开实施例中，确定模块102，用于将与第一数量无线接入网设备对应的第一数量模型训练结构输出层，确定为第一数量特有模型层。
在本公开实施例中，确定模块102，还用于将与第一数量无线接入网设备对应的模型训练结构输入层和隐藏层，确定为共享模型层，并获取多个无线接入网设备的数据，对数据添加与每个无线接入网设备对应的数据标识。对带有数据标识的所有数据进行分类处理，得到模型训练数据和模型标签值。将模型训练数据作为第一输入数据，输入至共享模型层，得到共享模型层输出的第一输出数据。向多个无线接入网设备发送模型标签值和第一输出数据。
在本公开实施例中，装置还包括：更新模块104。
更新模块104，用于响应于接收到多个无线接入网设备发送的训练损失值，基于训练损失值更新共享模型层的结构参数。
在本公开实施例中，更新模块104，用于对训练损失值进行加权，得到加权损失值。确定共享模型层当前的模型参数和模型学习率。基于加权损失值、模型参数和模型学习率，确定共享模型层的更新参数，并基于更新参数对共享模型层的结构参数进行更新。
在本公开实施例中，更新模块104，还用于响应于第T次更新共享模型层的结构参数，确定共享模型层训练完成，将第T次更新后的共享模型层的模型结构参数，发送至多个无线接入网设备组中的每个无线接入网设备。其中，T为预设更新共享模型层和特有模型层的次数，共享模型层结构参数用于无线接入网设备合成无线接入网设备订阅的模型。
在本公开实施例中，分组模块101，用于确定每个模型订阅请求包括的订阅模型的类型。基于类型，对模型订阅请求进行分组，得到第一组数的模型订阅请求。对无线接入网设备进行分组，得到第一组数的无线接入网设备组。
在本公开实施例中，更新模块104，还用于响应于存在新增的无线接入网设备，且新增的无线接入网设备满足模型训练条件，将与新增无线接入网设备对应的特有模型层的结构参数，发送至新增的无线接入网设备。或，响应于存在退出的无线接入网设备，重新确定第一数量模型训练结构。
图31是根据一示例性实施例示出的一种模型训练装置框图。参照图31，该装置200 应用于无线接入网设备，包括接收模块201。
接收模块201，用于接收OAM发送的特有模型层的结构参数。其中，特有模型层为OAM划分第一数量模型训练结构确定的。第一数量模型训练结构为OAM基于无线接入网设备组包括的第一数量无线接入网设备的模型订阅请求确定的。
在本公开实施例中，接收模块201，还用于接收OAM发送的模型标签值和第一输出数据。将第一输出数据作为特有模型层的输入，输入至特有模型层，得到特有模型层输出的第二输出数据。基于模型标签值和第二输出数据，确定训练损失值，并向OAM发送训练损失值。
在本公开实施例中，装置还包括：确定模块202。
确定模块202，用于基于模型训练数据携带的标识，在模型标签值中，确定与无线接入网设备对应的模型标签值。运算第二输出数据与训练标签值，确定训练损失值，并基于训练损失值对特有模型层的结构参数进行更新。
在本公开实施例中，接收模块201，还用于接收OAM发送的共享模型层的结构参数。基于模型共享层的结构参数和第T次更新后的特有模型层的结构参数确定订阅模型的结构参数。其中，T为预设更新共享模型层和特有模型层的次数。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
图32是根据一示例性实施例示出的一种用于模型训练的装置300的框图。例如，装置300可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等。
参照图32，装置300可以包括以下一个或多个组件：处理组件302，存储器304，电力组件306，多媒体组件308，音频组件310，输入/输出(I/O)接口312，传感器组件314，以及通信组件316。
处理组件302通常控制装置300的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件302可以包括一个或多个处理器320来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件302可以包括一个或多个模块，便于处理组件302和其他组件之间的交互。例如，处理组件302可以包括多媒体模块，以方便多媒体组件308和处理组件302之间的交互。
存储器304被配置为存储各种类型的数据以支持在装置300的操作。这些数据的示例包括用于在装置300上操作的任何应用程序或方法的指令，联系人数据，电话簿数据，消息，图片，视频等。存储器304可以由任何类型的易失性或非易失性存储设备或者它们的 组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电力组件306为装置300的各种组件提供电力。电力组件306可以包括电源管理系统，一个或多个电源，及其他与为装置300生成、管理和分配电力相关联的组件。
多媒体组件308包括在所述装置300和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一些实施例中，多媒体组件308包括一个前置摄像头和/或后置摄像头。当装置300处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件310被配置为输出和/或输入音频信号。例如，音频组件310包括一个麦克风(MIC)，当装置300处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器304或经由通信组件316发送。在一些实施例中，音频组件310还包括一个扬声器，用于输出音频信号。
I/O接口312为处理组件302和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件314包括一个或多个传感器，用于为装置300提供各个方面的状态评估。例如，传感器组件314可以检测到装置300的打开/关闭状态，组件的相对定位，例如所述组件为装置300的显示器和小键盘，传感器组件314还可以检测装置300或装置300一个组件的位置改变，用户与装置300接触的存在或不存在，装置300方位或加速/减速和装置300的温度变化。传感器组件314可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件314还可以包括光传感器，如CMOS或CCD图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件314还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件316被配置为便于装置300和其他设备之间有线或无线方式的通信。装置300可以接入基于通信标准的无线网络，如WiFi，2G或3G，或它们的组合。在一个示例性实施例中，通信组件316经由广播信道接收来自外部广播管理系统的广播信号或广播相关信 息。在一个示例性实施例中，所述通信组件316还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，装置300可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种包括指令的非临时性计算机可读存储介质，例如包括指令的存储器304，上述指令可由装置300的处理器320执行以完成上述方法。例如，所述非临时性计算机可读存储介质可以是ROM、随机存取存储器(RAM)、CD-ROM、磁带、软盘和光数据存储设备等。
图33是根据一示例性实施例示出的一种用于模型训练的装置400的框图。例如，装置400可以被提供为一服务器。参照图33，装置400包括处理组件422，其进一步包括一个或多个处理器，以及由存储器432所代表的存储器资源，用于存储可由处理组件422的执行的指令，例如应用程序。存储器432中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件422被配置为执行指令，以执行上述方法。
装置400还可以包括一个电源组件426被配置为执行装置400的电源管理，一个有线或无线网络接口450被配置为将装置400连接到网络，和一个输入输出(I/O)接口458。装置400可以操作基于存储在存储器432的操作系统，例如Windows ServerTM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTM或类似。
进一步可以理解的是，本公开中“多个”是指两个或两个以上，其它量词与之类似。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。字符“/”一般表示前后关联对象是一种“或”的关系。单数形式的“一种”、“所述”和“该”也旨在包括多数形式，除非上下文清楚地表示其他含义。
进一步可以理解的是，术语“第一”、“第二”等用于描述各种信息，但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开，并不表示特定的顺序或者重要程度。实际上，“第一”、“第二”等表述完全可以互换使用。例如，在不脱离本公开范围的情况下，第一信息也可以被称为第二信息，类似地，第二信息也可以被称为第一信息。
进一步可以理解的是，本公开实施例中尽管在附图中以特定的顺序描述操作，但是不应将其理解为要求按照所示的特定顺序或是串行顺序来执行这些操作，或是要求执行全部 所示的操作以得到期望的结果。在特定环境中，多任务和并行处理可能是有利的。
本领域技术人员在考虑说明书及实践这里公开的发明后，将容易想到本公开的其它实施方案。本申请旨在涵盖本公开的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本公开的一般性原理并包括本公开未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的，本公开的真正范围和精神由下面的权利要求指出。
应当理解的是，本公开并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本公开的范围仅由所附的权利要求来限制。

Claims (17)

  1. 一种模型训练方法，其特征在于，应用于操作维护管理OAM实体，所述方法包括：
    对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组，所述无线接入网设备组包括第一数量无线接入网设备；
    确定所述第一数量无线接入网设备对应的第一数量模型训练结构，并基于所述第一数量模型训练结构，确定第一数量特有模型层；
    向所述第一数量无线接入网设备，发送所述第一数量特有模型层的结构参数。
  2. 根据权利要求1所述的模型训练方法，其特征在于，所述确定所述第一数量无线接入网设备对应的第一数量模型训练结构，包括：
    确定所述第一数量无线接入网设备发送的第一数量模型订阅请求，并确定所述第一数量模型订阅请求的模型训练任务特性，所述模型训练任务特性用于指示模型层数和节点数；
    基于所述模型训练任务特性所指示的模型层数和节点数，确定第一数量的模型训练结构。
  3. 根据权利要求1或2所述的模型训练方法，其特征在于，所述基于所述第一数量模型训练结构，确定第一数量特有模型层，包括：
    将与第一数量无线接入网设备对应的第一数量模型训练结构输出层，确定为第一数量特有模型层。
  4. 根据权利要求1所述的模型训练方法，其特征在于，所述方法还包括：
    将与第一数量无线接入网设备对应的模型训练结构输入层和隐藏层，确定为共享模型层，并获取所述多个无线接入网设备的数据，对所述数据添加与每个无线接入网设备对应的数据标识；
    对带有所述数据标识的所有数据进行分类处理，得到模型训练数据和模型标签值；
    将所述模型训练数据作为第一输入数据，输入至所述共享模型层，得到所述共享模型层输出的第一输出数据；
    向所述多个无线接入网设备发送所述模型标签值和所述第一输出数据。
  5. 根据权利要求4所述的模型训练方法，其特征在于，所述方法还包括：
    响应于接收到所述多个无线接入网设备发送的训练损失值，基于所述训练损失值更新所述共享模型层的结构参数。
  6. 根据权利要求5所述的模型训练方法，其特征在于，所述基于所述训练损失值更 新所述共享模型层，包括：
    对所述训练损失值进行加权，得到加权损失值；
    确定所述共享模型层当前的模型参数和模型学习率；
    基于所述加权损失值、模型参数和模型学习率，确定所述共享模型层的更新参数，并基于所述更新参数对所述共享模型层的结构参数进行更新。
  7. 根据权利要求6所述的模型训练方法，其特征在于，所述基于所述更新参数对所述共享模型层进行更新之后，所述方法包括：
    响应于第T次更新所述共享模型层的结构参数，确定所述共享模型层训练完成，将第T次更新后的所述共享模型层的模型结构参数，发送至所述多个无线接入网设备组中的每个无线接入网设备；
    其中，所述T为预设更新共享模型层和特有模型层的次数，所述共享模型层结构参数用于无线接入网设备合成无线接入网设备订阅的模型。
  8. 根据权利要求1所述的模型训练方法，其特征在于，所述对发送模型订阅请求的无线接入网设备进行分组，得到至少一个无线接入网设备组，包括：
    确定每个所述模型订阅请求包括的订阅模型的类型；
    基于所述类型，对所述模型订阅请求进行分组，得到第一组数的模型订阅请求；
    对无线接入网设备进行分组，得到所述第一组数的无线接入网设备组。
  9. 根据权利要求1所述的模型训练方法，其特征在于，所述方法还包括：
    响应于存在新增的无线接入网设备，且所述新增的无线接入网设备满足模型训练条件，将与新增无线接入网设备对应的特有模型层的结构参数，发送至所述新增的无线接入网设备；
    或
    响应于存在退出的无线接入网设备，重新确定所述第一数量模型训练结构。
  10. 一种模型训练方法，其特征在于，应用于无线接入网设备，所述方法包括：
    接收OAM发送的特有模型层的结构参数；
    其中，所述特有模型层为OAM划分第一数量模型训练结构确定的；所述第一数量模型训练结构为OAM基于无线接入网设备组包括的第一数量无线接入网设备的模型订阅请求确定的。
  11. 根据权利要求10所述的模型训练方法，其特征在于，所述方法还包括：
    接收OAM发送的模型标签值和第一输出数据；
    将所述第一输出数据作为所述特有模型层的输入，输入至所述特有模型层，得到所述特有模型层输出的第二输出数据；
    基于所述模型标签值和所述第二输出数据，确定训练损失值，并向OAM发送所述训练损失值。
  12. 根据权利要求11所述的模型训练方法，其特征在于，所述基于所述模型训练数据和所述第二输出数据，确定训练损失值，包括：
    基于所述模型训练数据携带的标识，在所述模型标签值中，确定与所述无线接入网设备对应的模型标签值；
    运算所述第二输出数据与所述训练标签值，确定训练损失值，并基于所述训练损失值对所述特有模型层的结构参数进行更新。
  13. 根据权利要求10所述的模型训练方法，其特征在于，所述方法还包括：
    接收OAM发送的共享模型层的结构参数；
    基于所述模型共享层的结构参数和第T次更新后的特有模型层的结构参数确定订阅模型的结构参数；
    其中，所述T为预设更新共享模型层和特有模型层的次数。
  14. 一种模型训练装置，其特征在于，应用于操作维护管理OAM实体，所述装置包括：
    分组模块，用于对发送模型订阅请求的多个无线接入网设备进行分组，得到至少一个无线接入网设备组，所述无线接入网设备组包括第一数量无线接入网设备；
    确定模块，用于确定所述第一数量无线接入网设备对应的第一数量模型训练结构，并基于所述第一数量模型训练结构，确定第一数量特有模型层；
    发送模块，用于向所述第一数量无线接入网设备，发送所述第一数量特有模型层的结构参数。
  15. 一种模型训练装置，其特征在于，应用于无线接入网设备，所述装置包括：
    接收模块，用于接收OAM发送的特有模型层的结构参数；
    其中，所述特有模型层为OAM划分第一数量模型训练结构确定的；所述第一数量模型训练结构为OAM基于无线接入网设备组包括的第一数量无线接入网设备的模型订阅请求确定的。
  16. 一种模型训练装置，其特征在于，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为：执行权利要求1-9中任意一项所述的模型训练方法，或执行权利要求10-13中任意一项所述的模型训练方法。
  17. 一种非临时性计算机可读存储介质，当所述存储介质中的指令由移动终端的处理器执行时，使得移动终端能够执行权利要求1-9中任意一项所述的模型训练方法，或使得移动终端能够执行权利要求10-13中任意一项所述的模型训练方法。
PCT/CN2021/098008 2021-06-02 2021-06-02 一种模型训练方法、模型训练装置及存储介质 WO2022252162A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2021/098008 WO2022252162A1 (zh) 2021-06-02 2021-06-02 一种模型训练方法、模型训练装置及存储介质
CN202180001782.9A CN115735214A (zh) 2021-06-02 2021-06-02 一种模型训练方法、模型训练装置及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/098008 WO2022252162A1 (zh) 2021-06-02 2021-06-02 一种模型训练方法、模型训练装置及存储介质

Publications (1)

Publication Number Publication Date
WO2022252162A1 true WO2022252162A1 (zh) 2022-12-08

Family

ID=84322718

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/098008 WO2022252162A1 (zh) 2021-06-02 2021-06-02 一种模型训练方法、模型训练装置及存储介质

Country Status (2)

Country Link
CN (1) CN115735214A (zh)
WO (1) WO2022252162A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612153A (zh) * 2019-02-22 2020-09-01 华为技术有限公司 训练模型的方法和装置
CN112396070A (zh) * 2019-08-13 2021-02-23 中兴通讯股份有限公司 一种模型训练方法、装置和系统，以及预测方法和装置
US20210097381A1 (en) * 2019-09-27 2021-04-01 Canon Medical Systems Corporation Model training method and apparatus

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612153A (zh) * 2019-02-22 2020-09-01 华为技术有限公司 训练模型的方法和装置
CN112396070A (zh) * 2019-08-13 2021-02-23 中兴通讯股份有限公司 一种模型训练方法、装置和系统，以及预测方法和装置
US20210097381A1 (en) * 2019-09-27 2021-04-01 Canon Medical Systems Corporation Model training method and apparatus

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
QUALCOMM INC.: "Model Training Procedure", 3GPP TSG-RAN WG3 MEETING #112-E R3-211755, 6 May 2021 (2021-05-06), XP052002038 *

Also Published As

Publication number Publication date
CN115735214A (zh) 2023-03-03

Similar Documents

Publication Publication Date Title
JP6862586B1 (ja) ニューラルネットワークのトレーニング方法、装置および記憶媒体
CN109829433B (zh) 人脸图像识别方法、装置、电子设备及存储介质
CN107133094B (zh) 应用管理方法、移动终端及计算机可读存储介质
CN108304758B (zh) 人脸特征点跟踪方法及装置
TWI780751B (zh) 網路訓練方法、目標檢測方法、電子設備和電腦可讀儲存介質
CN111160448B (zh) 一种图像分类模型的训练方法及装置
US10970523B2 (en) Terminal and server for providing video call service
CN105830080A (zh) 使用特定于应用和特定于应用类型的模型进行移动设备行为的高效分类的方法和系统
CN107995370B (zh) 通话控制方法、装置及存储介质和移动终端
EP3923202A1 (en) Method and device for data processing, and storage medium
EP3671441B1 (en) Application management method and apparatus, storage medium, and electronic device
CN112926310B (zh) 一种关键词提取方法及装置
JP2020511055A (ja) アプリケーション制御方法およびアプリケーション制御装置
CN114097259A (zh) 一种通信处理方法、通信处理装置及存储介质
CN114071627A (zh) 网络切换方法、智能终端及存储介质
WO2021245327A1 (en) Collaborative machine learning
TW202133030A (zh) 圖像處理方法及圖像處理裝置、電子設備和電腦可讀儲存媒體
CN109217909B (zh) 天线切换方法、移动终端和计算机存储介质
CN112116095A (zh) 一种多任务学习模型训练的方法及相关装置
CN114025215A (zh) 文件处理方法、移动终端及存储介质
WO2022252162A1 (zh) 一种模型训练方法、模型训练装置及存储介质
CN117332844A (zh) 对抗样本生成方法、相关装置及存储介质
CN107480773B (zh) 训练卷积神经网络模型的方法、装置及存储介质
CN115146633A (zh) 一种关键词识别方法、装置、电子设备及存储介质
CN110659726B (zh) 一种图像处理方法、装置、电子设备及存储介质

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: 21943529

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 18566626

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21943529

Country of ref document: EP

Kind code of ref document: A1