Model training method and device, medical signal identification system and equipment and medium

Abstract

The invention relates to a model training method and device, a medical signal recognition system and equipment and a medium, and the model training method is applied to a server and comprises the steps: obtaining global model parameters; wherein the global model parameters are used for local training of the client to obtain local model parameters; determining aggregation weights of the local model parameters according to data distribution characteristics of the local model parameters obtained by training of the client in the local model parameters to be aggregated; and performing parameter aggregation according to the local model parameters of the client and the aggregation weight to obtain target global model parameters. Thus, the local model parameters for the aggregation model can be adaptively adjusted according to the aggregation weight, the influence of data distribution imbalance on model training is reduced, and the accuracy of the target global model obtained based on the target global model parameters in practical application is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular, to a model training method, device, medical signal recognition system, equipment and medium. Background Art

[0002] Medical signals are usually used to diagnose and monitor different diseases and physiological states, and can also be used to predict the development trend of diseases. The diagnosis and recognition of medical signals depends largely on the experience and skills of doctors. Intelligent auxiliary recognition of medical signals can effectively solve the current problems of uneven distribution of medical resources, high misdiagnosis rate, and backward diagnostic methods.

[0003] The relevant technology can perform medical signal recognition tasks by constructing a deep neural network model. However, since the medical signal data owned by each medical institution is small in scale and the data distribution is often uneven, the constructed deep neural network model has a low accuracy rate. Summary of the invention

[0004] The purpose of the present disclosure is to provide a model training method, apparatus, medical signal recognition system, equipment and medium, which combine federated learning and deep neural network models, and adaptively adjust local model parameters used for aggregation models according to aggregation weights, reduce the impact of uneven data distribution on model training, and improve the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0005] In order to achieve the above-mentioned purpose, according to a first aspect of an embodiment of the present disclosure, a model training method is provided, which is applied to a server, and the method includes:

[0006] Obtaining global model parameters; wherein the global model parameters are used by the client to obtain local model parameters through local training;

[0007] Determine the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained by client training in the local model parameters to be aggregated;

[0008] Parameter aggregation is performed based on the client's local model parameters and aggregation weights to obtain the target global model parameters.

[0009] Optionally, determining the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated includes:

[0010] Determine multiple clustering categories and the central parameter of each clustering category according to the similarity between the local model parameters obtained by training of multiple clients; wherein each clustering category corresponds to the local model parameter of at least one client;

[0011] The aggregation weight of the local model parameter is determined according to the difference between the local model parameter of each client and the central parameter of the cluster category corresponding to the local model parameter.

[0012] Optionally, the method further comprises:

[0013] Get the prediction accuracy of the client during local training;

[0014] The step of determining the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated includes:

[0015] The aggregation weight of the local model parameters is determined according to the data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated and the prediction accuracy.

[0016] Optionally, the server includes multiple sub-servers, and different sub-servers are used to aggregate local model parameters of clients with different training capabilities;

[0017] The method further comprises:

[0018] Assign corresponding sub-servers to target clients with the same training capabilities;

[0019] The step of performing parameter aggregation according to the local model parameters and aggregation weights of the client to obtain the target global model parameters includes:

[0020] In the sub-server corresponding to the target client, parameter aggregation is performed on the local model parameters and aggregation weights of the target client to obtain sub-global model parameters;

[0021] Determining an adaptive aggregation weight of a sub-server corresponding to the target client according to the prediction accuracy of the target client during local training;

[0022] Parameter aggregation is performed according to the sub-global model parameters and the adaptive aggregation weights to obtain the target global model parameters.

[0023] Optionally, the server includes multiple sub-servers, and different sub-servers are used to aggregate local model parameters of clients with different training capabilities;

[0024] The method further comprises:

[0025] Obtaining configuration information of the client; wherein the configuration information of the client is used to determine the training capability of the client;

[0026] Determining the number of local iterative trainings of the client in one round of iterative training according to the training capability of the client;

[0027] Obtaining the training duration of each round of iterative training of the client based on the local training performed according to the number of local iterative training times;

[0028] According to the training duration of each round of iterative training of the client, the training capability of the client and the number of local iterative training times are updated.

[0029] Optionally, updating the training capability of the client and the number of local iterative training times according to the training duration of each round of iterative training of the client includes:

[0030] Updating the training capability of the client according to the relationship between the training duration of each round of iterative training of the client and the preset training duration;

[0031] When the training capability of the client changes, updating the number of local iterative training times of the client;

[0032] The preset training duration is determined based on a reference duration and an evaluation coefficient, and the reference duration is determined based on the training duration of each round of iterative training of a target client having the same training capability.

[0033] According to a second aspect of an embodiment of the present disclosure, a model training method is provided, which is applied to a client, and the method includes:

[0034] Get the global model parameters sent by the server;

[0035] Performing local training according to the global model parameters and local training data to obtain local model parameters;

[0036] Among them, the local model parameters are used by the server to determine the aggregation weights corresponding to the local model parameters according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, and to perform parameter aggregation according to the local model parameters and the aggregation weights to obtain the target global model parameters.

[0037] Optionally, the method further comprises:

[0038] Determine the parameter responsiveness of the global model parameters obtained in the previous round of iterative training to the local model parameters of the client aggregated and used in the previous round of iterative training;

[0039] Determining the responsiveness loss of the previous round of iterative training according to the parameter responsiveness;

[0040] Determine the loss function of this round of iterative training based on the loss function used in the previous round of iterative training and the responsiveness loss;

[0041] The loss function of the current round of iterative training is used to adjust the local model parameters of the current round obtained by the current round of iterative training to obtain the local model parameters.

[0042] According to a third aspect of an embodiment of the present disclosure, a medical signal recognition system is provided, the system comprising:

[0043] A first acquisition module, used to acquire medical signal data to be identified;

[0044] A recognition module, used for inputting the medical signal data into a federated learning model to obtain a recognition result output by the federated learning model;

[0045] The federated learning model is determined by target global model parameters obtained through training based on the model training method described in any one of the first aspect or any one of the second aspect.

[0046] According to a fourth aspect of an embodiment of the present disclosure, a model training device is provided, which is applied to a server, and the device includes:

[0047] A second acquisition module is used to acquire global model parameters; wherein the global model parameters are used by the client to obtain local model parameters through local training;

[0048] A determination module, used to determine the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained by client training in the local model parameters to be aggregated;

[0049] The aggregation module is used to perform parameter aggregation based on the client's local model parameters and aggregation weights to obtain the target global model parameters.

[0050] According to a fifth aspect of an embodiment of the present disclosure, a model training device is provided, which is applied to a client, and the device includes:

[0051] The third acquisition module is used to obtain the global model parameters sent by the server;

[0052] A training module, used for performing local training according to the global model parameters and local training data to obtain local model parameters;

[0053] Among them, the local model parameters are used by the server to determine the aggregation weights corresponding to the local model parameters according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, and to perform parameter aggregation according to the local model parameters and the aggregation weights to obtain the target global model parameters.

[0054] According to a sixth aspect of an embodiment of the present disclosure, a non-temporary computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in any one of the first aspect or any one of the second aspect are implemented.

[0055] According to a seventh aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0056] a memory having a computer program stored thereon;

[0057] A processor is used to execute the computer program in the memory to implement the steps of the method described in any one of the first aspect or any one of the second aspect above.

[0058] Through the above technical solution, federated learning and deep neural network models are combined, and the corresponding aggregation weights are determined according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, so that the local model parameters used for the aggregation model can be adaptively adjusted according to the aggregation weights, thereby reducing the impact of uneven data distribution on model training and improving the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0059] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:

[0061] Figure 1 It is a flowchart of a model training method shown in an exemplary embodiment of the present disclosure.

[0062] Figure 2 It is a flowchart of a model training method shown in another exemplary embodiment of the present disclosure.

[0063] Figure 3 It is a schematic diagram of a model framework shown in an exemplary embodiment of the present disclosure.

[0064] Figure 4 The flowchart of determining the training capability of a client is shown in an exemplary embodiment of the present disclosure.

[0065] Figure 5 It is a flowchart of a federated learning training shown in an exemplary embodiment of the present disclosure.

[0066] Figure 6It is a block diagram of a medical signal recognition system shown in an exemplary embodiment of the present disclosure.

[0067] Figure 7 It is a block diagram of a model training device shown in an exemplary embodiment of the present disclosure.

[0068] Figure 8 It is a block diagram of a model training device shown in another exemplary embodiment of the present disclosure.

[0069] Fig. 9 It is a block diagram of an electronic device shown in an exemplary embodiment of the present disclosure.

[0070] Fig.10 is a block diagram of an electronic device shown in another exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0071] The specific implementation of the present disclosure is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the present disclosure, and is not used to limit the present disclosure.

[0072] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization given by the owner of the corresponding device.

[0073] As mentioned in the background technology, the accuracy of the deep neural network model constructed by the relevant technology is not high. Therefore, considering the problem of small data scale and overcoming the difficulties of data silos and privacy protection, federated learning and deep neural network models can be combined to collaborate in deep neural network model training. Due to the large differences in medical signal data among medical institutions, all medical signal data from different medical institutions are often unevenly distributed, which leads to uneven training and performance degradation of ordinary federated deep neural network models.

[0074] In view of this, the embodiments of the present disclosure provide a model training method, which combines federated learning and a deep neural network model, and determines the corresponding aggregation weight according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, so that the local model parameters used for the aggregation model can be adaptively adjusted according to the aggregation weight, thereby reducing the impact of uneven data distribution on model training and improving the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0075] It is understandable that the training process of the federated deep neural network model can be completed by the server and the client together, and the specific training process may include multiple rounds of iterative training. Among them, each round of iterative process involves the client obtaining local model parameters through multiple local iterative training, and the server aggregating the local model parameters obtained by all client training to obtain global model parameters.

[0076] Through the federated learning training model, data distributed in different participants can be integrated together to build a more comprehensive and representative global model, which can better adapt to various model usage scenarios. Participants can refer to clients with different training data.

[0077] For example, the participants can be client devices of various medical institutions, and the training data can be medical signal data of various medical institutions. On this basis, the server can collect local model parameters trained by each client based on its local medical signal data, and aggregate them to obtain global model parameters, thereby obtaining a global model with better medical signal recognition effect for different situations.

[0078] Figure 1 FIG. 1 is a flow chart of a model training method shown in an exemplary embodiment of the present disclosure. Figure 1 As shown, the model training method can be applied to a server and may include steps S101 to S103.

[0079] In step S101, global model parameters are obtained.

[0080] Among them, the global model parameters are used by the client to obtain local model parameters through local training.

[0081] It should be noted that in the first round of iterative training, the global model parameters may be randomly generated or may be a pre-trained model trained based on relevant medical signal data collected in advance. After the first round of training, the global model parameters may be aggregated in the previous round of iterative training.

[0082] In a possible implementation, for each round of iterative training, the server sends the global model parameters to each client, and the client applies the received global model parameters to the local model trained locally, and performs local training on the local model to obtain the local model parameters. On this basis, the server can obtain the local model parameters trained locally by each client.

[0083] In step S102, the aggregation weight of the local model parameters is determined according to the data distribution characteristics of the local model parameters obtained by client training in the local model parameters to be aggregated.

[0084] The local model parameters to be aggregated may refer to the local model parameters obtained by the client training participating in the federated training. The data distribution characteristics may include the central tendency characteristics representing the degree of difference between the data and the cluster center, the discrete characteristics representing the degree of data dispersion, and the frequency distribution characteristics representing the relative frequency of data occurrence. Different aggregation weights may be determined according to different data distribution characteristics.

[0085] For example, the data distribution feature is a discrete feature, thereby determining the discrete features of the local model parameters of the client in the local model parameters to be aggregated, and determining the aggregation weight of the local model parameters according to the degree of discreteness represented by the discrete features. When the degree of discreteness of the local model parameters is higher, it can be indicated that the distribution of the local model parameters is uneven, thereby assigning a smaller aggregation weight to the local model parameters to reduce the influence of the local model parameters on parameter aggregation. Correspondingly, when the degree of discreteness of the local model parameters is lower, it can be indicated that the distribution of the local model parameters is more balanced, thereby assigning a larger aggregation weight to the local model parameters to enhance the influence of the local model parameters on parameter aggregation.

[0086] In step S103, parameter aggregation is performed according to the local model parameters and aggregation weights of the client to obtain target global model parameters.

[0087] It can be understood that the aggregation weight is determined according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated. Therefore, the local model parameters used for the aggregation model can be reasonably and adaptively adjusted according to the aggregation weight, thereby reducing the impact of uneven data distribution on model training and improving the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0088] In a possible implementation, the data distribution feature is a central tendency feature that characterizes the degree of difference between the data and the cluster center. On this basis, the above step S102 may include:

[0089] Determine multiple clustering categories and the central parameter of each clustering category according to the similarity between the local model parameters obtained by training multiple clients; wherein each clustering category corresponds to the local model parameter of at least one client;

[0090] The aggregation weight of the local model parameters is determined according to the difference between the local model parameters of each client and the central parameter of the cluster category corresponding to the local model parameters.

[0091] It should be noted that the central tendency feature characterizing the difference between the local model parameters and the cluster center can be determined based on a clustering algorithm, wherein the clustering algorithm can include, for example, a density-based clustering algorithm (DBSCAN) and a K-means clustering algorithm (K-means).

[0092] For example, the local model parameters to be aggregated are classified by a clustering algorithm, and the clustering categories and the central parameters of each clustering category on the data distribution are determined, so that the central tendency characteristics that characterize the degree of difference between the local model parameters and the central parameters of the corresponding clustering categories can be determined, and the aggregation weight of the local model parameters is determined according to the central tendency characteristics. When the degree of difference between the local model parameters and the central parameters between the corresponding clustering categories is greater, it can be indicated that the distribution of the local model parameters is uneven, and thus a smaller aggregation weight can be assigned to the local model parameters to reduce the influence of the local model parameters on parameter aggregation. Correspondingly, when the degree of difference between the local model parameters and the central parameters between the corresponding clustering categories is smaller, it can be indicated that the distribution of the local model parameters is more balanced, and thus a larger aggregation weight can be assigned to the local model parameters to enhance the influence of the local model parameters on parameter aggregation.

[0093] Among them, the process of classifying the local model parameters to be aggregated by the clustering algorithm involves calculating the similarity between different objects, and the similarity can represent the degree of similarity between different objects. It should be understood that the greater the similarity between two objects, the smaller the difference between the two objects, and the smaller the similarity between the two objects, the greater the difference between the two objects. The objects involved in the similarity calculation may include the local model parameters and the center parameters of the clustering categories determined by clustering.

[0094] For example, the similarity calculation may be performed using the following formula 1.

[0095]

[0096] Among them, x i and x j They represent different objects involved in similarity calculation. and They represent the local model parameters obtained by different clients in the kth round of iterative training, and z represents the total number of iterative training times.

[0097] In the disclosed embodiment, taking the K-means clustering algorithm as an example, on the basis of randomly initializing q central parameters, for each local model parameter, its similarity with all central parameters can be compared, and each local model parameter is assigned to the cluster category corresponding to the central parameter with the greatest similarity, and then the central parameter of each cluster category is recalculated. On this basis, the steps of comparing the similarity between the local model parameters and the central parameters to recalculating the central parameters are repeated multiple times until the stopping condition is met. The value of q can be determined according to the actual situation, for example, it can be determined according to the total number of clients. The stopping condition can be determined according to the actual situation, for example, the central parameter no longer changes significantly or reaches a predetermined number of iterations.

[0098] It can be understood that the above formula 1 is a similarity calculation based on cosine similarity, and other similarity calculation methods may also be used in specific applications, such as Euclidean distance and Manhattan distance.

[0099] In addition, the local model parameters obtained through training of all clients may be clustered and subsequent operations may be continued, or all local model parameters to be aggregated may be divided into batches, and the local model parameters of each batch may be clustered and subsequent operations may be continued, which is not specifically limited in the present disclosure.

[0100] In this way, the impact of data bias and data imbalance on the aggregation model can be balanced through aggregation weights, thereby improving model accuracy and generalization.

[0101] In a possible implementation, the prediction accuracy of the client during local training can be used as a standard for the aggregation weight. Therefore, based on obtaining the prediction accuracy of the client during local training, the above step S102 may include:

[0102] The aggregation weight of the local model parameters is determined according to the data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated and the prediction accuracy.

[0103] For example, the following formula 2 may be used to determine the prediction accuracy of the client during local training.

[0104]

[0105] Among them, F a Represents the prediction accuracy, P represents the precision of the local model trained locally by the client, and R represents the recall of the local model trained locally by the client.

[0106] In a possible implementation, the following formula 3 is used to determine the aggregation weight of the local model parameters.

[0107]

[0108] Among them, m represents the total number of clients, k represents the kth round of iterative training, ω i represents the local model parameters involved in similarity calculation, c represents the central parameter of the clustering category corresponding to the local model parameters, D(ω i ,c) represents ω i The similarity between and c, θ represents the aggregation weight.

[0109] On this basis, parameter aggregation can be performed according to the local model parameters and the aggregation weights of the local model parameters to obtain the target global model parameters.

[0110] For example, parameter aggregation is performed according to the following formula 4.

[0111]

[0112] Among them, m represents the total number of clients. represents the local model parameters of the i-th client in this round of iterative training, θ i represents the aggregate weight of the local model parameters, Represents the target global model parameters obtained by aggregating local model parameters.

[0113] Therefore, the local model parameters used for the aggregation model can be adaptively adjusted according to the aggregation weight, reducing the impact of uneven data distribution on model training and improving the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0114] In one possible implementation, the parameter responsiveness of the global model to the local training data of the client can be determined locally on the client, and the responsiveness loss can be determined based on this, so that the local model parameters obtained in each round of local training can be updated according to the responsiveness loss. In this way, the impact of the differences in training data between different clients on model aggregation can be further alleviated, and key information can be retained to the greatest extent to avoid overfitting of unimportant information.

[0115] For example, based on the global model parameters obtained in the previous round of iterative training and the local model parameters of the client used in the previous round of iterative training, convolution operations and integral operations are constructed to simulate and determine the parameter responsiveness of the global model. Taking the calculation of each convolution kernel as an example, based on the attribution idea, the weight responsiveness during the convolution kernel calculation is introduced to reflect the parameter responsiveness of the global model to each input data of the client. The parameter responsiveness of each parameter of each neural node to the input data can be approximately calculated using the trapezoidal method.

[0116] In the disclosed embodiment, the definition of parameter responsiveness is as shown in the following formula 5 and formula 6.

[0117]

[0118] Among them, F s represents the parameter responsiveness, ∈ represents the number of elements contained in each channel in the convolution operation, C represents the convolution operation, Represents the relationship between the sth feature subgraph and the local model parameters of the r-1 layer when the input is x′+Δx Perform convolution operation. To avoid overfitting, use the hyperparameter λ for adjustment. represents the sth parameter of the rth layer, x′ represents a connection baseline constructed based on the target global model parameters obtained in the previous round of aggregation based on the integral gradient method, τ represents the path control variable, Δx represents the increment of the path, and the specific value of the increment can be set based on experience.

[0119] On this basis, a parameter responsivity matrix as shown in the following formula 7 can be constructed.

[0120]

[0121] Among them, v represents the number of convolution kernels in the rth layer, Represents the responsiveness loss determined locally by the client in the kth iteration of training.

[0122] In a possible implementation, the client can determine the parameter responsiveness of the global model parameters obtained in the previous round of iterative training to the local model parameters of the client aggregated and used in the previous round of iterative training, and determine the responsiveness loss of the previous round of iterative training on this basis. Thus, the loss function used locally can be updated based on the responsiveness loss, and the local model parameters of the current round of iterative training obtained in the current round can be adjusted according to the updated loss function.

[0123] Therefore, in the disclosed embodiment, the local model parameters acquired by the server may be obtained by the client adjusting the local model parameters obtained in the current round of iterative training according to the parameter responsiveness of the global model parameters obtained in the previous round of iterative training.

[0124] For example, the loss function used in the client local training may be as shown in the following formula 8 and formula 9.

[0125]

[0126] in, Represents the loss function used by the client in the kth round of iterative training, that is, the loss function used in the previous round of iterative training. represents the responsiveness loss determined locally by the client in the kth round of iterative training, that is, the responsiveness loss in the previous round of iterative training. q represents the updated loss function, that is, the loss function of this round of iterative training. t represents the time parameter, which can represent the number of iterations, the time step, or a certain stage of the model training process according to the actual situation. T represents the time constant, which can be used with t to control the rate of time decay or to affect the change of regularization equations over time according to the actual situation. represents the operation of converting the convolution operation C into a vector, R represents the total number of layers of the local model, h represents the hth neural node, μ represents the local model response loss hyperparameter, and x h,p,r Represents the local input data of the client at the pth iteration during the kth round of iterative training.

[0127] On this basis, the local model parameters of this round of iterative training obtained by adjusting the loss function of this round of iterative training can be used to obtain the local model parameters.

[0128] For example, the client can locally determine the local model parameters by using the following formula 10.

[0129]

[0130] Among them, W t q Represents the local model parameters of this round of iterative training, Represents the adjusted local model parameters, η is a hyperparameter representing the learning rate, which is used to control the step size of each update and its value can be set according to actual conditions.

[0131] In this way, the client analyzes the relationship constraints between local data and global model parameters based on the situation of its local data, and introduces its own weight parameter penalty, that is, responsiveness loss, for each client, so that the local model parameters obtained in this round of iterative training can be adaptively adjusted, thereby effectively avoiding overfitting of unimportant information while retaining key information to the greatest extent.

[0132] In addition, it should be noted that the various clients participating in the training have differences in storage, computing power, communication capabilities, etc., that is, the training capabilities of each client are different. These differences make the local training time of each client different, and the device may be disconnected, resulting in problems such as long model training time and low accuracy.

[0133] In the disclosed embodiment, considering the different training capabilities of different clients, the clients are classified according to their training capabilities, and different local iteration times are assigned, so as to perform different self-iteration strategies locally on the clients. On this basis, multiple sub-servers are configured to aggregate the local model parameters of clients with different training capabilities.

[0134] In a possible implementation, the server includes multiple sub-servers, and different sub-servers are used to aggregate local model parameters of clients with different training capabilities. On this basis, the technical solution provided by the embodiment of the present disclosure may include:

[0135] Get the client's configuration information;

[0136] According to the training capability of the client, determine the number of local iterative training times of the client in one round of iterative training;

[0137] Obtain the training duration of each round of iterative training based on the local training performed by the client according to the number of local iterative training times;

[0138] Update the client's training capability and local iterative training times based on the training duration of each round of iterative training on the client.

[0139] The configuration information of the client may include GPU model, CPU model, memory capacity, storage device information, network connection speed, etc. On this basis, clients with higher graphics card and processor performance, larger memory, and faster network connection speed can be determined as high-performance clients with higher training capabilities. Correspondingly, clients with lower graphics card and processor performance, smaller memory, and slower network connection speed can be determined as lower-performance clients with lower training capabilities.

[0140] Therefore, the training capability of the client can be determined according to the configuration information of the client. The higher the training capability of the client, the faster the training speed is, so more local iterative training times can be allocated. Correspondingly, the lower the training capability of the client, the slower the training speed is, so fewer local iterative training times can be allocated. In this way, the model training time can be balanced, the model aggregation time can be avoided to be too long, and the model performance can be improved.

[0141] It should be understood that the training capability of the client determined based on the client's configuration information may not be accurate. In order to obtain a more accurate classification, the client's training capability and the number of local iterative training times allocated to the client can be updated based on the training duration of each round of iterative training based on the local training number of the client.

[0142] Optionally, according to the training duration of each round of iterative training of the client, the step of updating the training capability of the client and the number of local iterative training times may include:

[0143] Update the training capability of the client according to the relationship between the training duration of each round of iterative training of the client and the preset training duration;

[0144] When the training capability of the client changes, update the number of local iteration training times of the client;

[0145] The preset training duration is determined based on the reference duration and the evaluation coefficient. The reference duration is determined based on the training duration of each round of iterative training of the target client with the same training capability. The evaluation coefficient can be set based on actual conditions and experience.

[0146] For example, the reference duration may be determined according to the following formula 11.

[0147]

[0148] in, represents the reference duration of clients with the same training capability, m represents the total number of clients with the same training capability, k represents the kth round of iterative training, Represents the training duration of the i-th client in the k-th round of iterative training.

[0149] In a possible implementation, the preset training duration can be set to a value greater than the product of the reference duration and the evaluation coefficient. On this basis, the average training duration of the client can be determined according to the training duration of each round of iterative training of the client, and whether to update the training capability of the client can be determined according to the relationship between the average training duration of the client and the preset training duration.

[0150] For example, when the average training duration of the client is greater than a preset training duration, it is determined that the training capability of the client has changed, so that the training capability of the client can be updated according to a reference duration standard corresponding to clients with other training capabilities.

[0151] Optionally, the technical solution provided by the embodiment of the present disclosure may also include:

[0152] Assign corresponding sub-servers to target clients with the same training capabilities.

[0153] It is understandable that different sub-servers are used to aggregate local model parameters of clients with different training capabilities. Therefore, after different sub-servers complete the aggregation of local model parameters of corresponding clients, a step of aggregating the parameters aggregated by different sub-servers is also involved. On this basis, the above step S103 may include:

[0154] In the sub-server corresponding to the target client, the local model parameters and aggregation weights of the target client are aggregated to obtain sub-global model parameters;

[0155] Determine the adaptive aggregation weight of the sub-server corresponding to the target client according to the prediction accuracy of the target client during local training;

[0156] Parameter aggregation is performed according to the sub-global model parameters and the adaptive aggregation weights to obtain the target global model parameters.

[0157] Among them, the step of performing parameter aggregation on the local model parameters and aggregation weights of the target client in the sub-server corresponding to the target client to obtain the sub-global model parameters can refer to the content involved in the above formula 4. That is, the target global model parameters obtained by the above formula 4 are used as sub-global model parameters in the embodiment of the present disclosure, and on this basis, the target global model parameters are further aggregated to obtain the target global model parameters.

[0158] In the embodiment of the present disclosure, on the basis of respectively performing parameter aggregation on each sub-server to obtain the corresponding sub-global model parameters, the sub-global model parameters can be aggregated using adaptive aggregation weights according to the following formula 12 to obtain the target global model parameters.

[0159]

[0160] Where n represents the number of sub-servers. represents the average prediction accuracy of the j-th client with the same training capability during local training, Represents the adaptive aggregation weight determined according to the prediction accuracy of the target client during local training. represents the sub-global model parameters, Indicates that the target global model parameters are obtained by aggregating the sub-global model parameters.

[0161] Therefore, the comprehensive aggregation weights and adaptive aggregation weights are used to adjust the local model parameters, which reduces the impact of uneven data distribution on model training and improves the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0162] Figure 2 FIG. 1 is a flow chart of a model training method shown in another exemplary embodiment of the present disclosure. Figure 2 As shown, the model training method can be applied to the client and may include steps S201 to S202.

[0163] In step S201, the global model parameters sent by the server are obtained.

[0164] In step S202, local training is performed according to the global model parameters and the local training data to obtain local model parameters.

[0165] Among them, the local model parameters are used by the server to determine the aggregation weights corresponding to the local model parameters according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, and to perform parameter aggregation according to the local model parameters and the aggregation weights to obtain the target global model parameters.

[0166] Through the above technical solution, federated learning and deep neural network models are combined, and the corresponding aggregation weights are determined according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, so that the local model parameters used for the aggregation model can be adaptively adjusted according to the aggregation weights, thereby reducing the impact of uneven data distribution on model training and improving the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0167] In a possible implementation, the technical solution provided by the embodiment of the present disclosure may also include:

[0168] Determine the parameter responsiveness of the global model parameters obtained in the previous round of iterative training to the local model parameters of the clients aggregated and used in the previous round of iterative training;

[0169] Determine the responsiveness loss of the previous round of iterative training according to the parameter responsiveness;

[0170] Determine the loss function for this round of iterative training based on the loss function and responsiveness loss used in the previous round of iterative training;

[0171] The loss function of this round of iterative training is used to adjust the local model parameters of this round obtained by this round of iterative training to obtain the local model parameters.

[0172] In this way, the local model parameters obtained by local client training can be adaptively adjusted through parameter responsiveness, thereby further alleviating the impact of differences in training data between different clients on model aggregation, and retaining key information to the greatest extent, avoiding overfitting of unimportant information.

[0173] In addition, it is not difficult to understand that different clients have different training capabilities. Clients can be classified according to their training capabilities and assigned different local iteration times. Therefore, clients with different training capabilities can perform different self-iteration strategies locally, thereby balancing the overall training time of the model in federated learning, avoiding excessive time for model aggregation, and improving model performance.

[0174] Regarding the model training method on the client side in the above-mentioned embodiment, the specific implementation method of each step has been described in detail in the embodiment of the model training method on the server side, and will not be elaborated here.

[0175] Figure 3 is a schematic diagram of a model framework shown in an exemplary embodiment of the present disclosure. Figure 3 As shown, the model framework provided by the embodiment of the present disclosure may include a federated training module 301 , a pre-classification module 302 , and a client adaptive model update module 303 .

[0176] Among them, the federated training module 301 can be used to implement adaptive aggregation strategies and improve federated learning through multiple sub-servers, thereby reducing the impact of uneven data distribution on model aggregation. The pre-classification module 302 can be used to classify clients according to their training capabilities, and assign different local iteration training times to clients according to the classification results, so that clients can perform different self-iteration strategies locally, thereby balancing the overall training time of the model in federated learning, avoiding excessive time for model aggregation, and improving model performance. The client adaptive model update module 303 can be used to determine the parameter responsiveness of the global model to the client's local training data, and design a client adaptive loss function based on the parameter responsiveness, thereby updating the local model parameters obtained from each round of local training. In this way, the impact of the differences in training data between different clients on model aggregation can be further alleviated, and key information can be retained to the greatest extent to avoid overfitting of unimportant information.

[0177] Continue to refer to Figure 3 For example, the improved federated learning system proposed in the embodiment of the present disclosure may include N clients, 1 central server and 4 sub-servers. On this basis, all clients can be classified into four categories of slow, relatively slow, relatively fast and fast according to the configuration information of the client and the actual training situation through the pre-classification module 302. In addition, the offline client can be regarded as not participating in the federated training.

[0178] The process of classifying according to the client configuration information can refer to Figure 4 The process shown. Figure 4 FIG. 1 is a flow chart of determining the training capability of a client as shown in an exemplary embodiment of the present disclosure. Figure 4 As shown, the step of determining the training capability of the client may include steps S401 to S406.

[0179] In step S401, the central server sends a pre-classification notification to each client.

[0180] In step S402, each client sends its own configuration information and a task preparation completion mark to the central server.

[0181] In step S403, the central server regards the client that has not received a reply as offline and does not participate in this federated training task, and makes an allocation based on the configuration information of the received client to determine the preliminary classification of the client.

[0182] For example, the configuration information of the client is scored according to the preset rules, so that the clients with different scores are classified into high-performance configuration, medium-high-performance configuration, medium-low-performance configuration and low-performance configuration. The preset rules can be determined according to the actual situation, and the present disclosure does not make specific restrictions on this. Clients with different configurations have different training capabilities, and correspond to the fast, relatively fast, relatively slow and slow client classifications respectively. q , C rq , C rs , C s}.

[0183] In step S404, the central server sends the global model parameters of this federated training task and the pre-classified local iterative training times to the corresponding client.

[0184] The federated training task may be a task related to auxiliary identification of medical signals. The number of local iterative training times for pre-classification includes the number of local iterative training times for each category determined according to the preliminary classification of the client.

[0185] For example, corresponding to the task of auxiliary identification of medical signals, the training data used locally by each client may include, for example, medical imaging data, medical record data, test data, vital signs data, and health management data, etc. In actual application scenarios, any type of medical data can be classified, and its classification labels can be arbitrarily set according to actual needs, and the present disclosure does not make specific limitations on this. For example, classification and identification can be performed on electrocardiogram signals, and the classification labels used may include, for example, relevant labels for auxiliary diagnosis of arrhythmias: atrial fibrillation, ventricular fibrillation, tachycardia, bradycardia, etc., relevant labels for auxiliary diagnosis of ischemic heart disease: myocardial infarction, angina pectoris, etc., relevant labels for auxiliary diagnosis of cardiomyopathy: dilated cardiomyopathy, hypertrophic cardiomyopathy, etc., relevant labels for auxiliary diagnosis of congenital heart disease: atrial septal defect, ventricular septal defect, etc., and relevant labels for auxiliary diagnosis of other heart diseases: pericarditis, myocarditis, etc.

[0186] In step S405, each client creates a local model for local training according to the received global model parameters, performs local training according to the corresponding number of local iterative training times, and records the training duration of each round of iterative training.

[0187] On this basis, the client can send the training duration of each round of iterative training to the central server. It should be understood that the client can also send the local model parameters obtained by local training to the corresponding sub-server to perform model aggregation for this round of iterative training.

[0188] In step S406, the central server receives the training duration of each round of iterative training from each client, and updates the client's category and the number of local iterative training times according to the pre-classification strategy.

[0189] Among them, the pre-classification strategy can refer to the aforementioned step of updating the client's training capability and the number of local iterative training times according to the relationship between the training duration of each round of iterative training of the client and the preset training duration.

[0190] For example, refer to the above formula 11 to determine the reference duration corresponding to different client categories: And set the evaluation coefficients {μ 1 , μ 2 , μ 3 , μ 4}. The value range of the evaluation coefficient can be [0, 1.5]. On this basis, the preset training time corresponding to different client classifications can be determined.

[0191] Therefore, it can be determined whether to update the client classification according to the relationship between the average training time of each client and the preset training time, and the updated client classification is obtained. q , C rq , C rs , C s}.

[0192] It is understandable that clients of the same category can q , C rq , C rs , C s} propose their own self-iteration strategies, such as C s The number of iterations is 1, C q The number of iterations is 4. The specific iteration strategy can be adjusted according to the actual situation to achieve the same outer loop in different sub-servers to improve the model training efficiency and the overall model recognition accuracy.

[0193] Continue to use Figure 3 Example, where the process of federated learning training between the client and the server can refer to Figure 5 The process shown. Figure 5FIG. 1 is a flowchart of a federated learning training shown in an exemplary embodiment of the present disclosure. Figure 5 As shown, the steps of federated learning training may include steps S501 to S513.

[0194] In step S501, the central server divides the clients into corresponding categories according to the pre-classification, and creates sub-servers corresponding to the slow, relatively slow, relatively fast, and fast categories.

[0195] In step S502, for each type of client, the similarity between the local model parameters of every two clients in the same type is calculated.

[0196] In step S503, the evaluation index of each client between classes is calculated.

[0197] The evaluation index may include the aggregation weight and the adaptive aggregation weight in the aforementioned embodiment.

[0198] In step S504, the clients of each category are divided into different cluster categories according to the clustering algorithm and the evaluation index.

[0199] In step S505, different sub-servers adjust the number of local iterative training times of the client according to the self-iteration strategy of this category.

[0200] In step S506, each client performs local model training.

[0201] In step S507, each client uploads the trained local model parameters to the corresponding sub-server.

[0202] In step S508, the sub-server aggregates the local model parameters of the clients of the same category according to the adaptive aggregation strategy of the sub-server.

[0203] In step S509, it is determined whether the iteration of the sub-server is completed.

[0204] It can be understood that, when it is determined that the iteration of the sub-server is finished, the process proceeds to step S510. When it is determined that the iteration of the sub-server is not finished, the process returns to step S506.

[0205] In step S510, each sub-server uploads the sub-global model parameters obtained by aggregation of the category to the central server.

[0206] In step S511, the central server aggregates the sub-global model parameters according to the central adaptive aggregation strategy.

[0207] In step S512, it is determined whether the total iterative training is completed.

[0208] It can be understood that, when it is determined that the total iterative training is finished, the process proceeds to step S513. When it is determined that the total iterative training is not finished, the process returns to step S506.

[0209] In step S513, target global model parameters are obtained and sent to each client.

[0210] In this way, by implementing the above-mentioned model training method through the model framework provided by the embodiment of the present disclosure, the model training efficiency and model performance can be effectively improved, thereby improving the accuracy of the target global model corresponding to the target global model parameters obtained based on model training in practical applications.

[0211] The specific implementation methods of the above steps have been described in detail in the embodiments of the relevant model training method and will not be elaborated here.

[0212] Based on the same inventive concept, an embodiment of the present disclosure provides a medical signal recognition system. Figure 6 is a block diagram of a medical signal recognition system 600 shown in an exemplary embodiment of the present disclosure. Figure 6 The medical signal recognition system 600 includes a first acquisition module 601 and a recognition module 602 .

[0213] The first acquisition module 601 is used to acquire medical signal data to be identified.

[0214] The recognition module 602 is used to input the medical signal data into the federated learning model to obtain the recognition result output by the federated learning model.

[0215] Among them, the federated learning model is determined by the target global model parameters obtained by training based on any of the above-mentioned model training methods.

[0216] Through the above technical solution, federated learning and deep neural network models are combined, and the corresponding aggregation weights are determined according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, so that the local model parameters used for the aggregation model can be adaptively adjusted according to the aggregation weights, thereby reducing the impact of uneven data distribution on model training and improving the accuracy of the target global model obtained based on the target global model parameters in practical applications.

[0217] Regarding the medical signal recognition system 600 in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the model training method, and will not be elaborated here.

[0218] Based on the same inventive concept, an embodiment of the present disclosure provides a model training device. Figure 7is a block diagram of a model training device 700 shown in an exemplary embodiment of the present disclosure. Figure 7 The model training device 700 may include a second acquisition module 701, a determination module 702 and an aggregation module 703.

[0219] The second acquisition module 701 is used to acquire global model parameters.

[0220] Among them, the global model parameters are used by the client to obtain local model parameters through local training.

[0221] The determination module 702 is used to determine the aggregation weight of the local model parameters according to the data distribution characteristics of the local model parameters obtained by client training in the local model parameters to be aggregated.

[0222] Aggregation module 703 is used to perform parameter aggregation according to the local model parameters of the client and the aggregation weight to obtain the target global model parameters.

[0223] In one embodiment, the determination module 702 is used to:

[0224] Determine multiple clustering categories and the central parameter of each clustering category according to the similarity between the local model parameters obtained by training multiple clients; wherein each clustering category corresponds to the local model parameter of at least one client;

[0225] The aggregation weight of the local model parameters is determined according to the difference between the local model parameters of each client and the central parameter of the cluster category corresponding to the local model parameters.

[0226] In one embodiment, the second acquisition module 701 is further used to: obtain the prediction accuracy of the client during local training.

[0227] On this basis, the determination module 702 is used to:

[0228] The aggregation weight of the local model parameters is determined according to the data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated and the prediction accuracy.

[0229] In one embodiment, the server includes multiple sub-servers, and different sub-servers are used to aggregate local model parameters of clients with different training capabilities. On this basis, the model training device 700 also includes a distribution module, which is used to:

[0230] Assign corresponding sub-servers to target clients with the same training capabilities.

[0231] Aggregation module 703 is used to:

[0232] In the sub-server corresponding to the target client, the local model parameters and aggregation weights of the target client are aggregated to obtain sub-global model parameters;

[0233] Determine the adaptive aggregation weight of the sub-server corresponding to the target client according to the prediction accuracy of the target client during local training;

[0234] Parameter aggregation is performed according to the sub-global model parameters and the adaptive aggregation weights to obtain the target global model parameters.

[0235] In one embodiment, the model training device 700 further includes an updating module, which is used to:

[0236] Obtaining configuration information of the client; wherein the configuration information of the client is used to determine the training capability of the client;

[0237] According to the training capability of the client, determine the number of local iterative training times of the client in one round of iterative training;

[0238] Obtain the training duration of each round of iterative training based on the local training performed by the client according to the number of local iterative training times;

[0239] Update the client's training capability and local iterative training times based on the training duration of each round of iterative training on the client.

[0240] In one embodiment, the update module is used to:

[0241] Update the training capability of the client according to the relationship between the training duration of each round of iterative training of the client and the preset training duration;

[0242] When the training capability of the client changes, update the number of local iteration training times of the client;

[0243] The preset training duration is determined based on the reference duration and the evaluation coefficient, and the reference duration is determined based on the training duration of each round of iterative training of a target client with the same training capability.

[0244] Regarding the model training device 700 in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the relevant model training method, and will not be elaborated here.

[0245] Based on the same inventive concept, an embodiment of the present disclosure provides a model training device. Figure 8 8 is a block diagram of a model training device 800 shown in another exemplary embodiment of the present disclosure. Figure 8 , the model training device 800 may include a third acquisition module 801 and a training module 802.

[0246] The third acquisition module 801 is used to acquire the global model parameters sent by the server.

[0247] The training module 802 is used to perform local training according to the global model parameters and local training data to obtain local model parameters.

[0248] Among them, the local model parameters are used by the server to determine the aggregation weights corresponding to the local model parameters according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, and to perform parameter aggregation according to the local model parameters and the aggregation weights to obtain the target global model parameters.

[0249] In one embodiment, the model training device 800 further includes an adjustment module, which is used to:

[0250] Determine the parameter responsiveness of the global model parameters obtained in the previous round of iterative training to the local model parameters of the client aggregated and used in the previous round of iterative training;

[0251] Determine the responsiveness loss of the previous round of iterative training according to the parameter responsiveness;

[0252] Determine the loss function for this round of iterative training based on the loss function and responsiveness loss used in the previous round of iterative training;

[0253] Among them, the loss function of this round of iterative training is used to adjust the local model parameters of this round obtained by this round of iterative training to obtain the local model parameters.

[0254] Regarding the model training device 800 in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the relevant model training method, and will not be elaborated here.

[0255] Fig. 9 FIG. 9 is a block diagram of an electronic device 900 according to an exemplary embodiment. Fig. 9 As shown, the electronic device 900 may be provided as a client, including: a processor 901 and a memory 902. The electronic device 900 may also include one or more of a multimedia component 903, an input / output (I / O) interface 904, and a communication component 905.

[0256] Among them, the processor 901 is used to control the overall operation of the electronic device 900 to complete all or part of the steps in the model training method on the client side. The memory 902 is used to store various types of data to support the operation of the electronic device 900, which may include, for example, instructions for any application or method used to operate on the electronic device 900, and application-related data, such as contact data, messages sent and received, pictures, audio, video, etc. The memory 902 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, disk or optical disk. The multimedia component 903 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 902 or sent through the communication component 905. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 904 provides an interface between the processor 901 and other interface modules, and the other interface modules may be keyboards, mice, buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 905 is used for wired or wireless communication between the electronic device 900 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them, is not limited here. Therefore, the corresponding communication component 905 may include: Wi-Fi module, Bluetooth module, NFC module, etc.

[0257] In an exemplary embodiment, the electronic device 900 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned client-side model training method.

[0258] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, and when the program instructions are executed by the processor, the steps of the above-mentioned client-side model training method are implemented. For example, the computer-readable storage medium can be the above-mentioned memory 902 including program instructions, and the above-mentioned program instructions can be executed by the processor 901 of the electronic device 900 to complete the above-mentioned client-side model training method.

[0259] Fig.10 1 is a block diagram of an electronic device 1000 according to an exemplary embodiment. For example, the electronic device 1000 may be provided as a server. Fig.10 The electronic device 1000 includes a processor 1022, which may be one or more, and a memory 1032 for storing a computer program executable by the processor 1022. The computer program stored in the memory 1032 may include one or more modules, each corresponding to a set of instructions. In addition, the processor 1022 may be configured to execute the computer program to perform the above-mentioned server-side model training method.

[0260] In addition, the electronic device 1000 may further include a power supply component 1026 and a communication component 1050, wherein the power supply component 1026 may be configured to perform power management of the electronic device 1000, and the communication component 1050 may be configured to implement communication, for example, wired or wireless communication, of the electronic device 1000. In addition, the electronic device 1000 may further include an input / output (I / O) interface 1058. The electronic device 1000 may operate based on an operating system stored in the memory 1032.

[0261] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, and when the program instructions are executed by the processor, the steps of the above-mentioned server-side model training method are implemented. For example, the non-transitory computer-readable storage medium can be the above-mentioned memory 1032 including program instructions, and the above-mentioned program instructions can be executed by the processor 1022 of the electronic device 1000 to complete the above-mentioned server-side model training method.

[0262] In another exemplary embodiment, a computer program product is also provided, which includes a computer program that can be executed by a programmable device, and the computer program has a code portion for executing the above-mentioned model training method when executed by the programmable device.

[0263] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings; however, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, a variety of simple modifications can be made to the technical solution of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0264] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present disclosure will not further describe various possible combinations.

[0265] In addition, various embodiments of the present disclosure may be arbitrarily combined, and as long as they do not violate the concept of the present disclosure, they should also be regarded as the contents disclosed by the present disclosure.

Claims

1. A model training method, characterized in that: Applied to a server, the method comprises: Obtaining global model parameters; wherein the global model parameters are used by the client to obtain local model parameters through local training; Determine the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained by client training in the local model parameters to be aggregated; Parameter aggregation is performed based on the client's local model parameters and aggregation weights to obtain the target global model parameters.

2. The model training method according to claim 1, characterized in that: The step of determining the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated includes: Determine multiple clustering categories and the central parameter of each clustering category according to the similarity between the local model parameters obtained by training multiple clients; wherein each clustering category corresponds to the local model parameter of at least one client; The aggregation weight of the local model parameter is determined according to the difference between the local model parameter of each client and the central parameter of the cluster category corresponding to the local model parameter.

3. The model training method according to claim 1, characterized in that: The method further comprises: Get the prediction accuracy of the client during local training; The step of determining the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated includes: The aggregation weight of the local model parameters is determined according to the data distribution characteristics of the local model parameters obtained through client training in the local model parameters to be aggregated and the prediction accuracy.

4. The model training method according to claim 3, characterized in that: The server includes a plurality of sub-servers, and different sub-servers are used to aggregate local model parameters of clients with different training capabilities; The method further comprises: Assign corresponding sub-servers to target clients with the same training capabilities; The step of performing parameter aggregation according to the local model parameters and aggregation weights of the client to obtain the target global model parameters includes: In the sub-server corresponding to the target client, parameter aggregation is performed on the local model parameters and aggregation weights of the target client to obtain sub-global model parameters; Determining an adaptive aggregation weight of a sub-server corresponding to the target client according to the prediction accuracy of the target client during local training; Parameter aggregation is performed according to the sub-global model parameters and the adaptive aggregation weights to obtain the target global model parameters.

5. The model training method according to claim 1, characterized in that: The server includes a plurality of sub-servers, and different sub-servers are used to aggregate local model parameters of clients with different training capabilities; The method further comprises: Obtaining configuration information of the client; wherein the configuration information of the client is used to determine the training capability of the client; Determining the number of local iterative trainings of the client in one round of iterative training according to the training capability of the client; Obtaining the training duration of each round of iterative training of the client based on the local training performed according to the number of local iterative training times; According to the training duration of each round of iterative training of the client, the training capability of the client and the number of local iterative training times are updated.

6. The model training method according to claim 5, characterized in that: The updating of the training capability of the client and the number of local iterative training times according to the training duration of each round of iterative training of the client includes: Updating the training capability of the client according to the relationship between the training duration of each round of iterative training of the client and the preset training duration; When the training capability of the client changes, updating the number of local iterative training times of the client; The preset training duration is determined based on a reference duration and an evaluation coefficient, and the reference duration is determined based on the training duration of each round of iterative training of a target client having the same training capability.

7. A model training method, characterized in that: Applied to a client, the method comprises: Get the global model parameters sent by the server; Performing local training according to the global model parameters and local training data to obtain local model parameters; Among them, the local model parameters are used by the server to determine the aggregation weights corresponding to the local model parameters according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, and to perform parameter aggregation according to the local model parameters and the aggregation weights to obtain the target global model parameters.

8. The model training method according to claim 7, characterized in that: The method further comprises: Determine the parameter responsiveness of the global model parameters obtained in the previous round of iterative training to the local model parameters of the client aggregated and used in the previous round of iterative training; Determining the responsiveness loss of the previous round of iterative training according to the parameter responsiveness; Determine the loss function of this round of iterative training based on the loss function used in the previous round of iterative training and the responsiveness loss; The loss function of the current round of iterative training is used to adjust the local model parameters of the current round obtained by the current round of iterative training to obtain the local model parameters.

9. A medical signal recognition system, characterized in that: The system comprises: A first acquisition module, used to acquire medical signal data to be identified; A recognition module, used for inputting the medical signal data into a federated learning model to obtain a recognition result output by the federated learning model; The federated learning model is determined by target global model parameters obtained by training according to the model training method described in any one of claims 1 to 6 or any one of claims 7 to 8.

10. A model training device, characterized in that: Applied to a server, the device comprises: A second acquisition module is used to acquire global model parameters; wherein the global model parameters are used by the client to obtain local model parameters through local training; A determination module, used to determine the aggregation weight of the local model parameters according to data distribution characteristics of the local model parameters obtained by client training in the local model parameters to be aggregated; The aggregation module is used to perform parameter aggregation based on the client's local model parameters and aggregation weights to obtain the target global model parameters.

11. A model training device, characterized in that: Applied to a client, the device comprises: The third acquisition module is used to obtain the global model parameters sent by the server; A training module, used for performing local training according to the global model parameters and local training data to obtain local model parameters; Among them, the local model parameters are used by the server to determine the aggregation weights corresponding to the local model parameters according to the data distribution characteristics of the local model parameters in the local model parameters to be aggregated, and to perform parameter aggregation according to the local model parameters and the aggregation weights to obtain the target global model parameters.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 or any one of claims 7 to 8 are implemented.

13. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method described in any one of claims 1 to 6 or any one of claims 7 to 8.

Images