A surface material classification method based on asynchronous federated learning and DDQN
By employing asynchronous federated learning and the DDQN method, the problems of uneven data distribution and differences in computing power in federated learning systems are solved, achieving efficient surface material classification in an asynchronous environment while ensuring data security and model accuracy.
Patent Information
- Application Number
- CN202310597443.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2043-05-25
AI Technical Summary
Existing federated learning systems suffer from uneven data distribution, differences in computing power, and privacy restrictions during data synchronization and model training across multiple clients, resulting in insufficient model training accuracy.
Asynchronous federated learning and DDQN methods are adopted. The client is randomly selected by the server for training. ResNet50 is used to process visual and tactile data, a DDQN environment is constructed, and state, action and reward functions are set. The model parameters are updated by combining stochastic gradient descent, and the parameters are aggregated on the server. A new scheduling algorithm is used to deal with the problems of uneven data and poor communication.
Without sharing data, asynchronous training reduces communication, improves model accuracy and robustness, adapts to differences in computing power and availability among different clients, and ensures data security.
Smart Images
Figure CN116630775B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of remote operation, and particularly relates to a surface material classification method based on asynchronous federated learning and DDQN. BACKGROUND
[0002] With the development of machine learning, federated learning as a new type of distributed machine learning framework meets the model training of multiple clients under the premise of data security. In the model training process, only model parameters are exchanged between the server and the client, and each client does not need to upload any original data. In the actual federated learning scene, although the addition of multiple clients brings more data, it also increases the difficulty of client synchronization, such as uneven data distribution. In practical applications, data is distributed in multiple devices, and the data in each device is limited in size and diversity, which cannot train an accurate DDQN for the entire distributed data group locally; on the other hand, due to privacy restrictions, data cannot be centrally shared, and these problems often affect the accuracy of model training, so a distributed DDQN algorithm is needed to train a DDQN representing the entire population.
[0003] The federated learning system is usually composed of a server and a client, and its architecture is similar to that of a parameter server. The client trains a local model using private data. The server aggregates the model parameters of the client and updates the global model parameters. Existing research usually uses a synchronous federated learning method to achieve this, which often only considers ideal situations and ignores special cases. When there are many clients and data, it is very difficult to achieve global synchronization due to the different computing capabilities, availability and completion times of each client.
[0004] Therefore, in view of the above status, it is urgent to develop a surface material classification method based on asynchronous federated learning and DDQN to overcome the deficiencies in current practical applications. SUMMARY
[0005] In view of the deficiencies in the prior art, the purpose of the embodiments of the present application is to provide a surface material classification method based on asynchronous federated learning and DDQN to solve the problems in the background art.
[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:
[0007] A surface material classification method based on asynchronous federated learning and DDQN, comprising the following steps:
[0008] Step 1, the server establishes a global DDQN model and initializes the model parameters w0, and starts the client scheduling and model parameter aggregation processes in parallel;
[0009] Step two, the server triggers training in the client scheduling process randomly selects part of the client, and sends the global model parameters and the communication round τ to the client;
[0010] Step three, the client processes the visual touch data using ResNet50, builds a DDQN environment, and sets the state, action and reward function;
[0011] Step four, calculate the time difference error, and update the model parameters through stochastic gradient descent;
[0012] Step five, the client updates the parameters through K times of parameter update, and the parameters are transmitted to the server, and the server updates the global parameters using the aggregation algorithm in the parameter aggregation process;
[0013] Step six, the server selects the client for training again, completes a communication, and designs a scheduling aggregation algorithm for the complex scene of poor communication and uneven data distribution.
[0014] As a further technical solution of the application, in step one, the specific way of establishing a global DDQN model and initializing the model parameter w0 by the server is:
[0015] The server defines the evaluation network Q and the target network Q' of DDQN, and the two networks are constructed identically, both of which are composed of 5 fully connected layers, and the neuron numbers of the 5 fully connected layers are 4096, 1024, 512, 128 and 108, and the activation functions of the fully connected layers are all ReLU;
[0016]
[0017] The server initializes its network parameters to w0, establishes a client scheduling and parameter aggregation process for subsequent model training, and establishes a queue queue to store the subsequent client model parameters and the binary tuple
[0018] As a further technical solution of the application, in step two, the specific method of the server in the client scheduling process is:
[0019] In each round of communication, the server selects n clients from N clients in the client scheduling process to trigger the training task, and selects the clients in an equal probability manner, and issues the parameter training, that is, the probability of each client being selected is
[0020] The server sends the current global model parameters and communication round information to the selected client, and the selected client receives the server global model parameters and communication round (w t= w0,t = 1), update the local model parameters w τ k and participate in the communication round tau:
[0021]
[0022] The client trains the local DDQN classification model using the above model parameters.
[0023] As a further technical solution of the application, in step three, the specific way for the client to build the DDQN environment is:
[0024] The client converts the local tactile acceleration data into a frequency spectrum graph using short-time Fourier transform with a Hamming window, and inputs it into the ResNet50 network. The ResNet50 network performs dimension reduction processing on the input visual image and tactile frequency spectrum graph, obtaining a visual feature L-dimensional row vector A tactile feature L-dimensional row vector And the sum of the two is a 2L-dimensional visual-tactile fusion row vector
[0025] The client sets the state, action, and reward functions, wherein the state s is the visual-tactile fusion row vector The action a is to assign a label to a state, and the action space a is defined as {1,...,M}; The reward r measures whether the action is correctly performed, which depends on the true class label. Taking the mth class of material as an example;
[0026]
[0027] The client combines the state, action, and reward functions, and uses the parameters sent by the server to establish a Q(s, a; w τ k network.
[0028] As a further technical solution of the application, in step four, the specific way for the client to update the parameters through stochastic gradient descent is:
[0029] The client updates the model parameters w τ k according to the received model parameters w τ k ;
[0030] Then the target network and the evaluation network are Q(s, a, w(t-t mod c)); Q(s, a, w(t)), respectively.
[0031] Calculate a time difference TD target y t :
[0032] y t = r t + gamma * Q(s t+1 , a ★ ; w(t-t mod c))
[0033]
[0034] With the TD target information, the evaluation network establishes a loss function:
[0035]
[0036] Where s t , a t , y t are randomly extracted from the experience pool, and the loss function is used to update the parameters of the evaluation network:
[0037]
[0038] Where alpha is the learning rate;
[0039] Thus, one random gradient descent update of the model is completed, and after c gradient descent updates, the network parameters are synchronized to the target network.
[0040] As a further technical solution of the application, in step five, the specific way of aggregating the server model parameters is:
[0041] After K training of the client, K is a small integer value, the number of times of participating in communication tau = tau + 1, and the model parameters w τ k and the binary tuple (w τ k , tau) are placed in the queue queue.
[0042] When the queue queue is not empty, the server takes out the parameters in the queue, and performs parameter aggregation in the model parameter aggregation process, and the aggregation method is:
[0043]
[0044] Where s (tau, t) is the staleness function, representing the lag between the number of times the client participates in training and the total number of communications, tau represents the number of times the client participates in parameter update, and t represents the total number of times the server updates the parameters;
[0045] After completing the parameter aggregation, the global model parameters of the server are updated to w t , which is obtained by the client in the scheduling process.
[0046] As a further technical solution of the present application, in step six, the specific way of the server selecting the client again is:
[0047] After the parameter aggregation, the server obtains new global parameters, detects the number of the queue queue (w τ k ,τ), when the number of the binary tuple is less than a certain constant, the server again schedules according to the scheduling algorithm, and the global model parameters w t are transmitted to the client to trigger the corresponding client to train and take a new scheduling algorithm:
[0048]
[0049] Where p(k) is the probability of the client being selected, and in the case of poor communication of individual clients, the weight of the client in parameter aggregation is adjusted;
[0050] A new aggregation algorithm is adopted to cope with the uneven distribution of local data of the client:
[0051] The new aggregation algorithm takes the client with the most material sample quantity as the benchmark, and the number of material types owned by each client is compared with the number of material types owned by the benchmark client to define the uneven rate σ k of the client. k Mapping σ * k :
[0052]
[0053] Then the new parameter aggregation formula is:
[0054] α t ←α×s(τ,t)×σ * k
[0055] w t ←α t w k +(1-α t )w t-1 ;
[0056] When the local data is unevenly distributed, the above parameter aggregation formula is used, and after multiple communications, a robust classifier for classifying surface materials is obtained.
[0057] As a further technical solution of the present application,
[0058] Compared with the prior art, the present application has the beneficial effects that:
[0059] The application can not only better represent and summarize the global model by using a large amount of training data from different clients, but also can train a DDQN capable of representing the entire population without sharing data, so that each user can train a DDQN model representing the entire data source, which not only significantly reduces the amount of communication data, but also ensures the data security of each device, and most importantly, the asynchronous training method is widely used in traditional distributed stochastic gradient descent, which can effectively reduce the influence of model inaccuracy caused by uneven data distribution and different computing power, availability and completion time of each client.
[0060] In order to make the structure characteristics and effects of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS
[0061] Fig. 1 The flowchart of the surface material classification method based on asynchronous federated learning and DDQN provided by the embodiment of the present application.
[0062] Fig. 2 The parameter transfer flowchart of the surface material classification method based on asynchronous federated learning and DDQN provided by the embodiment of the present application.
[0063] Fig. 3 The comparison chart of the verification accuracy of the delay function proposed by the present application and the traditional hinge delay function under the condition of poor communication and uneven data distribution. DETAILED DESCRIPTION
[0064] In order to make the structure characteristics and effects of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments.
[0065] The specific implementation of the present application will be described in detail below in combination with specific embodiments.
[0066] As shown in Figs. 1 to 3 A surface material classification method based on asynchronous federated learning and DDQN provided by an embodiment of the present application includes the following steps:
[0067] Step one, the server establishes a global DDQN model and initializes the model parameters w0, and starts the client scheduling and model parameter aggregation processes in parallel;
[0068] The specific way of the server establishing a global DDQN model and initializing the model parameters w0 is:
[0069] The server defines the evaluation network Q and the target network Q' of DDQN, both of which are constructed in the same way, i.e., by 5 fully connected layers with 4096, 1024, 512, 128 and 108 neurons respectively, and the activation function of each fully connected layer is ReLU;
[0070]
[0071] The server initializes its network parameters as w0, establishes a client scheduling and parameter aggregation process for subsequent model training, and establishes a queue queue to store subsequent client model parameters and the binary tuple composed of the communication round τ
[0072] Step two, the server randomly selects part of the client in the client scheduling process to trigger training, and sends the global model parameters and communication round τ to the client;
[0073] The specific method of the server in the client scheduling process is:
[0074] In each round of communication, the server selects n clients from N clients in the client scheduling process to trigger the training task. In order to ensure fairness and randomness, the client selection is performed in an equal probability manner, and the parameter training is issued, i.e., the probability of each client being selected is
[0075] The server sends the current global model parameters and communication round information to the selected client, and the selected client receives the server global model parameters and communication round (w t = w0, t = 1), updates the local model parameters w τ k and the communication round τ:
[0076]
[0077] The client trains the local DDQN classification model using the above model parameters;
[0078] Step three, the client processes the visual and tactile data using ResNet50, builds a DDQN environment, and sets the state, action and reward functions;
[0079] The specific approach of the client to build the DDQN environment is:
[0080] The client converts the local tactile acceleration data into a frequency spectrum graph using short-time Fourier transform with a Hamming window, and inputs it into the ResNet50 network. The ResNet50 network performs dimension reduction processing on the input visual image and tactile frequency spectrum graph to obtain a visual feature L-dimensional row vector L-dimensional tactile feature vector And the sum of the two is a 2L-dimensional visual-tactile fusion row vector
[0081] The client sets the state, action and reward function, where the state s is a visual-tactile fusion row vector The action a is to assign a label to a state, and the action space a is defined as {1,...,M}; The reward r measures whether the action is performed correctly, depending on the true class label, here taking the mth class material as an example;
[0082]
[0083] The client combines the state, action and reward function, and uses the parameters sent by the server To establish the Q(s,a;w τ k ) network;
[0084] Step four, calculate the time difference error, and update the model parameters by stochastic gradient descent;
[0085] The specific way for the client to update the parameters by stochastic gradient descent is:
[0086] The client updates the model parameters w τ k by K times of client local gradient descent, that is, w(t)←w τ k ;
[0087] Then the target network and the evaluation network are Q(s,a,w(t-t mod c)) and Q(s,a,w(t)), respectively.
[0088] Calculate a time difference TD target y t :
[0089] y t = r t + γQ(s t+1 ,a ★ ; w(t-t mod c))
[0090]
[0091] Using the TD target information, the evaluation network establishes a loss function:
[0092]
[0093] Where s t ,a t ,y tRandomly draw from the experience pool, update the parameters of the evaluation network using this loss function:
[0094]
[0095] Wherein, alpha is the learning rate;
[0096] Thus complete a random gradient descent update of the model, after experiencing c times of gradient descent, synchronize the network parameters to the target network;
[0097] Step five, the client updates the parameters Pass to the server, and the server updates the global parameters using the aggregation algorithm in the parameter aggregation process;
[0098] The specific way of server model parameter aggregation is:
[0099] After K times of training, K is a small integer value (if K value is too large, the number of client iterations is too much, which will make the model overfitting in local data), the number of times of participating in communication τ = τ + 1, the model parameters w τ k And the binary tuple (w τ k ,τ) is placed in the queue queue;
[0100] When the queue queue is not empty, the server takes out the parameters in the queue, and aggregates the parameters in the model parameter aggregation process, and the aggregation method is:
[0101]
[0102] Wherein, s(τ,t) is the staleness function, which represents the lag between the number of times the client participates in training and the total number of communications, τ represents the number of times the client participates in parameter update, and t represents the total number of times the server updates the parameters;
[0103] After completing the parameter aggregation, update the global model parameters of the server to w t , for the client to obtain in the scheduling process;
[0104] Step six, the server selects the client to train again, completes a communication, and designs a scheduling aggregation algorithm for complex scenarios of poor communication and uneven data distribution;
[0105] The specific way of server selecting client again is:
[0106] After parameter aggregation, the server obtains new global parameters, detects (w τ k, the number of two-tuples is less than a certain constant, the server again schedules according to the scheduling algorithm, and the global model parameters w t are transmitted to the client, triggering the corresponding client to train and take a new scheduling algorithm:
[0107]
[0108] where p(k) is the probability of being selected by the client, and in the case of poor communication of individual clients, the weight of the client in parameter aggregation is adjusted to reduce the impact on the classification effect of the model, so as to ensure the robustness of the model.
[0109] A new aggregation algorithm is adopted to deal with uneven distribution of client local data:
[0110] The new aggregation algorithm takes the client with the largest number of material samples as the reference, and the number of material types owned by each client is defined as the uneven rate of the client compared to the number of material types of the reference client k , when k σ k is taken as the aggregation weight, the variance is too large, which may cause the problem of ignoring small data clients, so * σ k is mapped to obtain the aggregation weight coefficient t :
[0111]
[0112] Then the new parameter aggregation formula is:
[0113] α t ←α×s(τ,t)×σ * k
[0114] w t ←α t w k +(1-α t )w t-1 ;
[0115] When the local data distribution is uneven, the parameter aggregation formula can be used to reduce the impact on the model and improve the performance of the model. After multiple communications, a robust classifier for classifying surface materials is obtained.
[0116] The applicability of the surface material classification method based on asynchronous federated learning and DDQN to material classification is analyzed through simulation experiment data. In the simulation experiment, the pytorch framework of the python software is used, and the specific content is as follows:
[0117] The simulation takes a python multithreaded simulation of the real environment, creates a scheduling thread, a parameter aggregation thread, a master thread, and N client threads, and the client thread is in a suspended state; use the pytorch framework of python to build DDQN evaluation network and target network, the two networks are constructed identically; they are composed of 5 fully connected layers, and the number of neurons of the 5 fully connected layers is 4096, 1024, 512, 128 and 108, and the activation function of the fully connected layer is ReLU.
[0118] Start the master thread to create the queue queue, and initialize the model parameters w = w0, and the communication round t = 1;
[0119] When the scheduling thread detects that the number of parameter, communication round two-tuple in the queue is less than the specified number, call the random function, use uniform distribution, randomly select n client threads to start;
[0120] Client thread: get the parameter w from the master thread, assign it to the DDQN network, and input the visual and tactile fusion feature data as the input state, define the reward function, calculate the time difference error, and use Adam stochastic gradient descent K times to update the client local parameter w τ k The number of times τ two-tuple of communication with the client is stored in the queue queue;
[0121] Parameter aggregation thread: when the thread runs, detect the queue, when the queue is not empty, the aggregation thread takes out the queue parameters, the number of times the client participates in communication, adopts the delay function, and aggregates the parameters according to the aggregation algorithm. Multiple threads work simultaneously to complete the training task.
[0122] The simulation data selects LTM108 database to evaluate the material classification method based on DDQN and federated learning, LTM108 database contains 108 materials, and each material has 20 pictures;
[0123] First, 2 pictures of each material are used for verification, and 2 pictures are used for testing; then the remaining 16 pictures are adjusted for random cropping of 4 sub-images; for each existing tactile sample, randomly crop 4 sliding sub-accelerations, that is, construct 64 groups of training data for each material; send the visual image sample of the material into the pre-trained ResNet50 model, and get a 2048-dimensional feature vector through the global average pooling layer; at the same time, the processed tactile acceleration spectrum of the material is also sent into the pre-trained ResNet50 model, and a 2048-dimensional feature vector is obtained from the global average pooling layer; finally, the above two 2048-dimensional feature vectors are spliced to construct a 4096-dimensional visual and tactile fusion feature vector; that is, a total of (108x64, 2048) data is obtained;
[0124] Six client threads are established to simulate real client devices, 108 kinds of material data are evenly distributed to the six client threads as local data, that is, each slave device obtains 18 kinds of material training data, which is a balanced scene; (46, 32, 16, 8, 4, 2) kinds of materials are respectively allocated to the client to simulate possible data imbalance scenarios of each client in actual operation; when the scheduling thread randomly selects the client thread to start, the probability of selecting a certain client thread is 1 / 20 of other devices to simulate the poor communication scenario; data imbalance and poor communication are both possible scenarios in actual system operation. These scenarios are simulated to test the robustness of the system;
[0125] The hyperparameters adopted in the simulation are as follows: the communication round T is 2000 rounds, the DDQN discount factor γ is 0.1, the learning rate is 0.0001, the batch size for training each time is 32, and the number of client iterations K for each communication is 3;
[0126] After 2000 communications, the trained network is saved, and the trained network is used to classify the test set to obtain the classification accuracy of the test set.
[0127] Table 1
[0128]
[0129] Table 1 is the result of federated learning + DDQN distributed architecture and DDQN centralized training on LMT108 data set
[0130] Table 2
[0131]
[0132] Table 2 is the result of the delay function and the traditional delay function hinge on the LMT108 data set under the condition of simulating poor communication and data distribution imbalance.
[0133] It can be seen from the analysis of the experimental results that the method proposed in the application has excellent classification ability, and from Table 2 and Fig. 3 It can be concluded that for the situation of poor communication, device disconnection and data distribution imbalance, the system still has good classification performance, the robustness of the system is guaranteed, and from Table 1, the classification performance of federated DDQN and centralized DDQN is almost the same, but the data security of the client of the system is effectively guaranteed, and the communication overhead is effectively reduced.
[0134] The above only describes the preferred embodiments of the application and should not be used to limit the application. Any modification, equivalent replacement and improvement made within the spirit and principle of the application should be included in the protection scope of the application.
Claims
1. A surface material classification method based on asynchronous federated learning and DDQN, characterized in that, Comprising the following steps: Step one, the server establishes a global DDQN model and initializes the model parameters The client scheduling and model parameter aggregation processes are started in parallel. Step two, the server randomly selects part of the client in the client scheduling process to trigger training, and sends the global model parameters and communication round to the client to the client; Step three, the client uses ResNet50 to process the visual and tactile data, builds a DDQN environment, sets the state, action and reward function; Step four, calculate the time difference error, update the model parameters through stochastic gradient descent; The specific way for the client to update the parameters through stochastic gradient descent is: The client performs , according to the received model parameters , a ; The target network and the evaluation network are respectively ; calculating a time difference TDtarget : ； Use TD target information to evaluate the network to establish a loss function: ； wherein From the experience pool, randomly draw, using this loss function, update the parameters of the evaluation network: ； wherein, is the learning rate; With this, a single stochastic gradient descent update of the model is completed, and the process is repeated After the second gradient descent, the network parameters are synchronized to the target network Step five, the client passes the updated parameters to the server, which performs a global parameter update using the aggregation algorithm in the parameter aggregation process. Step five, the client passes the updated parameters to the server, which performs a global parameter update using the aggregation algorithm in the parameter aggregation process. Step five, the client passes the updated parameters to the server, which performs a global parameter update using the aggregation algorithm in the parameter aggregation process. Step six, the server selects the client for training again, completes a communication, and designs a scheduling aggregation algorithm for complex scenes with poor communication and uneven data distribution; The specific way for the server to select the client again is: After parameter aggregation, the server obtains new global parameters and checks the queue. middle The number of pairs is determined by the scheduling algorithm. When the number of pairs is less than a certain constant, the server schedules again according to the scheduling algorithm, and the global model parameters are adjusted. This information is transmitted to the client, triggering the corresponding client to perform training and adopt the new scheduling algorithm. ； wherein, is the selected probability of the client, and in the case of poor communication of individual clients, the weight of the client in the parameter aggregation is adjusted; Take new aggregation algorithm to deal with uneven distribution of local data of client: The new aggregation algorithm uses the client with the most material samples as the benchmark. The ratio of the number of material types possessed by each client to the number of material types possessed by the benchmark client is defined as the non-uniformity rate of that client. ,right Perform mapping to obtain aggregate weight coefficients. : ； Then there is a new parameter aggregation formula: ； When the local data is unevenly distributed, the above parameter aggregation formula is used, and after multiple communications, a robust classifier for classifying surface materials is obtained.
2. The surface material classification method based on asynchronous federated learning and DDQN according to claim 1, characterized in that, In step one, the server establishes a global DDQN model and initializes the model parameters The specific approach is: The service end defines an evaluation network of DDQN and a target network , the two networks are constructed completely identically, and are composed of 5 fully connected layers, the neuron numbers of the 5 fully connected layers are 4096, 1024, 512, 128 and 108 respectively, and the activation functions of the fully connected layers are all ; ； The server initializes its network parameters as , establishes a client scheduling and parameter aggregation process for subsequent model training, establishes a queue to store subsequent client model parameters and participates in communication rounds composed of a two-tuple .
3. The surface material classification method based on asynchronous federated learning and DDQN according to claim 1, characterized in that, In step two, the specific method for the server to schedule the client process is: In each round of communication, the server selects from The server triggers the training task in the selected clients, and assigns the parameter training in an equal probability manner, i.e., the probability of each client being selected is The server sends the current global model parameters and the communication round information to the selected clients, and the selected clients receive the server global model parameters and the communication round , update the local model parameters and participate in the communication round : ； The client trains the local DDQN classification model using the above model parameters.
4. The surface material classification method based on asynchronous federated learning and DDQN according to claim 1, characterized in that, In step three, the specific way for the client to build the DDQN environment is: The client converts the local tactile acceleration data into a frequency spectrum graph by using a short-time Fourier transform with a Hamming window, and inputs the frequency spectrum graph into a ResNet50 network, the ResNet50 network performs dimension reduction processing on the input visual image and tactile frequency spectrum graph to obtain a visual feature L-dimensional row vector , a tactile feature L-dimensional row vector , and adds the two to obtain a 2L-dimensional visual-tactile fusion row vector ; The client sets the state, action, and reward function, where the state is a visual-tactile fusion row vector ; the action is to assign a label to a state, the action space is defined as ; the reward measures whether the action is performed correctly, depending on the true class label, with the first class material as an example; ； The client combines the state, action and reward function, using the parameters sent by the server , to establish a network.
5. The surface material classification method based on asynchronous federated learning and DDQN according to claim 1, characterized in that, In step five, the specific way for the server to aggregate the model parameters is: The client passes After the sub-training, For smaller integer values, the number of times Model parameters updated by gradient descent And the binary tuple Put into the queue In; When the queue If not empty, the server takes out the parameters in the queue, and carries out parameter aggregation in the model parameter aggregation process, and the aggregation method is: ； wherein, is a function representing the hysteresis between the round of client participation in training and the total number of communications, represents the number of times the client participates in parameter updates, represents the total number of times the server performs parameter updates; After completing parameter aggregation, the global model parameters of the update server are updated as for the clients to obtain in the scheduling process.