CN-121981297-A - Hierarchical federal learning method for sustainable deployment diagnosis of dynamic user clusters

CN121981297ACN 121981297 ACN121981297 ACN 121981297ACN-121981297-A

Abstract

The application discloses a hierarchical federation learning method for sustainable deployment diagnosis of a dynamic user cluster, which relates to the federation learning field, and comprises the steps that each local user performs sustainable training on a local model by adopting a loss function based on newly-added data in a local data set to obtain a trained local model; the method comprises the steps of obtaining a local data set, carrying out feature extraction on the local data set by each local user to obtain a feature matrix, clustering all local users by a central server based on the feature matrix of each local user to obtain a plurality of user clusters, determining the contribution weight of each local user in the user clusters by central service aiming at any user cluster, carrying out weighted average on the local model parameters trained by all local users in the user clusters based on the contribution weight to obtain global model parameters, and carrying out bearing fault diagnosis on each local user based on the global model parameters.

Inventors

QIN NA
DU JIAHAO
HUANG DEQING
XU JUNQI
DONG HAIRONG
CAI LI

Assignees

西南交通大学

Dates

Publication Date: 20260505
Application Date: 20251225

Claims (10)

1. A hierarchical federation learning method for dynamic user cluster sustainable deployment diagnostics, the method being applied to a hierarchical federation learning system comprising a central server and a plurality of local users, the method comprising: each local user acquires vibration signal data of a bearing to be diagnosed in real time to obtain a local data set, and based on newly-added data in the local data set, performs sustainable training on a local model by adopting a loss function to obtain a trained local model, wherein the loss function comprises a cross entropy loss term, a regularization term for preventing forgetting of knowledge and a regularization term for reducing characteristic distribution difference between the newly-added data and historical data; Each local user performs feature extraction on the local data set to obtain a feature matrix; The central server clusters all local users based on the feature matrix of each local user to obtain a plurality of user clusters; Aiming at any user cluster, the central server determines the contribution weight of each local user in the user cluster based on the local model parameters trained by each local user in the user cluster and the contribution value obtained by each local user by adopting a saproliferation interpretation method, and carries out weighted average on the local model parameters trained by all local users in the user cluster based on the contribution weight to obtain the cluster model parameters of the user cluster; And each local user updates a local model based on the global model parameters, and performs bearing fault diagnosis by adopting the updated local model.
2. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 1, wherein the expression of the loss function is: ; ; ; ; Wherein, the As a function of the total loss, Is the first A local multi-class cross entropy loss function for the newly added data, Is the first Group new data and the first EWC regularization terms between the set of historical data, Is the first Group new data and the first KLD regularization terms between the group history data, As a first super-parameter, the first super-parameter, As a second super-parameter, the second super-parameter, As a result of the third super-parameter, For the number of categories of the newly added data, For the category index of the newly added data, Is the first The real tag corresponding to the newly added data is like, Is the first The class newly added data is based on the predictive label of the activation function, To learn the first Group new data The values of the individual parameters are used to determine, Is the first Group history data below Parameters of Is used as a reference to the optimum value of (a), As a subscript of the parameters of the local model, As a total number of parameters of the local model, For the diagonal elements of the Fisher information matrix, Is the first The probability distribution of the newly added data of the group, Is the first The probability distribution of the group history data, As the total number of probability values in the probability distribution, Is used for indexing the probability value, Is the first The first data in the newly added group The probability value of the individual(s), Is the first Group history data of the first Probability values of the individuals.
3. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnosis according to claim 1, wherein the central server clusters all local users based on a feature matrix of each local user to obtain a plurality of user clusters, and specifically comprises: the center server determines a candidate range of the clustering quantity K, and initializes a Gaussian mixture model containing K Gaussian components aiming at any clustering quantity K in the candidate range, wherein each Gaussian component comprises a mixture coefficient, a mean vector and a covariance matrix; The central server adopts an expected maximization algorithm to iteratively optimize parameters of the Gaussian mixture model, wherein the iterative optimization comprises the steps of calculating posterior probability of each local user characteristic matrix belonging to each Gaussian component based on the current parameters of the Gaussian mixture model and the characteristic matrix of each local user, and updating a mixing coefficient, a mean vector and a covariance matrix of each Gaussian component based on the posterior probability; the center server calculates contour coefficients and variance ratio criteria for the optimized Gaussian mixture model under each clustering quantity K, and determines optimal clustering quantity based on the contour coefficients and the variance ratio criteria; and dividing each local user into a user cluster corresponding to the Gaussian component with the maximum posterior probability based on the optimal clustering number and the optimized Gaussian mixture model corresponding to the optimal clustering number by the central server.
4. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 3, wherein the posterior probability of the feature matrix of the kth user belonging to each gaussian component is determined using the following formula: ; Wherein, the The feature matrix for the kth user belongs to The posterior probability of the individual gaussian components, Is the first The mixing coefficients of the individual gaussian components, Feature matrix for kth user In the first place Probability density functions under the individual gaussian components, Is the first The mean vector of the individual gaussian components, Is the first The covariance matrix of the individual gaussian components, Is the first The mixing coefficients of the individual gaussian components, For the number of gaussian components, Is an index of the gaussian component, Feature matrix for kth user In the first place Probability density functions under the individual gaussian components, Is the first The mean vector of the individual gaussian components, Is the first Covariance matrix of the gaussian components.
5. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 4, wherein the mixing coefficients, mean vectors, and covariance matrices for each gaussian component are updated using the following formula: ; ; ; Wherein, the For the number of local users, For the index of the local user, Is the first Feature vectors of individual local users.
6. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 4, wherein the expression of the calculated profile coefficients is: ; Wherein, the For dissimilarity within a user cluster, For dissimilarity with the nearest neighbor user cluster, Is the contour coefficient value.
7. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 4, wherein the expression of variance ratio criteria is: ; Wherein, the Is the value of the variance ratio criterion, The trace being the sum of squares of the differences between classes, Is a trace of the sum of squares of the dispersion within the class, Is the number of local users.
8. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnosis according to claim 1, wherein the feature matrix of each local user comprises feature vectors of a plurality of samples, the feature vectors of each sample comprise a plurality of features, and the process of obtaining the contribution value by each local user by adopting a saprolitic additive interpretation method specifically comprises: Selecting partial samples from the feature vectors of a plurality of samples in the local user to form a background sample set aiming at any local user, and selecting the feature vector of at least one sample from the background sample set as a target sample; for any target sample, calculating the feature similarity weight between the target sample and each background sample in a background sample set; Based on the feature vector of each background sample, obtaining an original prediction result of each background sample by adopting the local model trained by the local user; For any feature of the feature vector of any background sample, replacing the feature with a value of a corresponding feature in the target sample to obtain a replaced feature vector of the background sample corresponding to the feature; Based on the replacement feature vector, a local model trained by the local user is adopted to obtain a corresponding replacement feature prediction result; obtaining marginal contribution of the feature on the background sample based on the original prediction result and the replacement feature prediction result; For each feature of the target sample, carrying out weighted summation on marginal contributions of the feature on all background samples and feature similarity weights of corresponding background samples respectively to obtain a saprolitic additive interpretation value of the feature on the target sample; And determining the contribution value of the local user based on the saprolitic additive interpretation values of all target samples of the local user.
9. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 1, wherein the feature similarity weight between the target sample and any background sample is determined and calculated using the following formula: ; Wherein, the In order to be a sample of the object, As a background sample, For feature similarity weights between the target sample and the background sample, For the number of features in each sample, To be in addition to the current characteristics All but the feature set(s), For the number of features contained in the feature set, For the j-th feature in the feature set, Is characterized by At the position of And (3) with The euclidean distance between them, Is the value of the j-th feature of the target sample, Is the value of the j-th feature of the background sample.
10. The hierarchical federal learning method for dynamic user cluster-oriented sustainable deployment diagnostics of claim 1, wherein the saprolitic additive interpretation of the features for the target sample is determined using the formula: ; Wherein, the For a saprolipram additive interpretation value of the feature for the target sample, As a set of background samples, As a background sample, For feature similarity weights between the target sample and the background sample, As an original prediction result of the background sample, In order to replace the result of the feature prediction, Is an alternative feature.

Description

Hierarchical federal learning method for sustainable deployment diagnosis of dynamic user clusters Technical Field The application relates to the field of federal learning, in particular to a hierarchical federal learning method for sustainable deployment diagnosis of a dynamic user cluster. Background In recent years, the popularity of artificial intelligence algorithms is continuously accelerated, wherein the deep learning algorithm successfully permeates into various fields such as fault diagnosis and the like by virtue of excellent accuracy and low requirements on field expertise, and shows strong application value and potential. With the increasing attention of data security, high-quality tag data cannot be directly used and is difficult to summarize and train in some industries, and meanwhile, data monopoly may exist among enterprises, so that collaborative sharing of data becomes difficult to be implemented. How to effectively mine information stored in data without invading the privacy of the data becomes a difficult problem to be solved urgently. For this reason, federal learning architecture is suitable for intelligent industrial scenarios with high data security requirements, aiming at solving the above-mentioned problems by multiparty data holder security collaborative modeling. The 'data island' dilemma existing in various fields can be broken through on the premise of protecting the privacy and safety of data through federal learning. However, most existing federal learning frameworks are designed for static settings, where participating clients and their local data sets remain fixed. In real-world industrial deployments, client participation is dynamic in nature due to changes in hardware performance, network conditions, and operating environment. New clients may join the federation at any time, while other clients may exit after their local model meets specified performance thresholds. At the same time, the continuously monitored sensors installed on each device may cause an increasing local data volume and a constant change in data heterogeneity between clients. The variability and diversity created by such dynamic client clusters has a significant impact on global model performance and aggregation efficiency, as follows: (1) The user's irregular joining and exiting can cause unpredictable interference to the overall federal model training, subject to the effects of actual hardware performance and network bandwidth. In particular, large data volume users have a great dominant role in the federal framework, and can have higher weights when model aggregation occurs, in which case their entry and exit can severely impact the performance of the current federal global model. In addition, there is great difference between the user individuals, and it is difficult to directly improve the model performance together through sharing data knowledge, and even negative effects can be brought, so that model loss can not be converged, and the global efficiency of federal learning is hindered. (2) The quality of the model of the user under the federal framework is uneven due to the influence of the data quality and the category distribution of the user, so that the contribution to the global model is different when the model is aggregated. If all users' contributions to the global model are considered equal, or weight fine-tuning is simply based on the amount of data, the global efficiency of federal learning is also reduced. In the federal scenario of dynamic user clusters, it is a better strategy to quantify the degree of influence of each user, either positively or negatively, on the global model, and further reset the user weights. (3) In the deployed federal learning framework, each user collects the latest data in an online manner, resulting in an ever-increasing data set that is difficult to fix. When the data is updated every time, if all the data are summarized and retrained according to the previous experience, the training efficiency is reduced, the calculation cost is increased greatly, and the training cost which is difficult to estimate can be generated at a later stage. This can pose serious challenges to the computational performance of the hardware, rendering the training process very lengthy and even in some cases impossible to complete within given time constraints, which is detrimental to the long-term efficient deployment of federal learning. From the background, how to optimize the federal learning framework when federal modeling diagnosis is implemented under the dynamic user cluster can be clearly obtained, model drift influence caused by user cluster variability and variability is avoided, and meanwhile, calculation cost is reduced, so that the sustainability of federal learning deployment in an industrial scene is ensured, and the problem to be solved urgently is solved. Disclosure of Invention The application aims to provide a hierarchical federal learning method for sustainable