A dynamic weighted personalized federated meta-learning fault diagnosis method based on unsupervised and grouping mechanism

Abstract

The application discloses a kind of dynamic weighting personalized federated meta-learning fault diagnosis methods based on unsupervised and grouping mechanism, to improve the accuracy and efficiency of mechanical fault diagnosis.The method first acquires the fault bearing vibration signal under different equipment, generates pseudo-label for the data set lacking label.Further, by designing a personalized federated meta-learning algorithm, clients with similar features are grouped to address the problem of non-independent and identically distributed data.In addition, the method introduces a dynamic weighting strategy to optimize the global model updating process.In this way, the method not only effectively utilizes distributed data, but also dynamically adjusts the learning strategy according to the data characteristics of each client, ultimately achieving more accurate and personalized fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of component fault diagnosis of industrial equipment, and particularly to a fault diagnosis method based on federated learning. BACKGROUND

[0002] Rotary mechanical equipment plays a vital role in the fields of aerospace, petrochemical industry, wind power, transportation, medical equipment, etc. The failure of key components such as bearings and gears will directly affect the safety of the system. Therefore, monitoring the operating state and identifying the occurring faults are of great significance to the normal operation of mechanical equipment.

[0003] With the development of artificial intelligence technology, intelligent diagnosis methods have become the "main force" of mechanical fault diagnosis. Although intelligent diagnosis methods have been successful, these methods need to use sufficient and comprehensive training data. However, in real industrial scenarios, considering economic and time factors, users usually have difficulty collecting sufficient training data. This situation restricts the application of intelligent diagnosis methods in real industries. An ideal choice is to directly aggregate multiple user data, which can greatly save the time cost of obtaining samples. However, in practice, due to industry competition, interest conflicts, privacy security, etc., there are insurmountable barriers between data sources. These data can only be mastered by the owner himself, and it is difficult to realize information exchange and data integration. Even in different departments of the same company, due to privacy security, data security or complex management processes, data cannot be integrated. Therefore, directly aggregating multiple user data is not feasible in real industries.

[0004] Federated learning technology that meets privacy protection and data security is one of the solutions to the above situation. The essence of federated learning is a distributed machine learning technology, and the feature of this technology is that users train models locally without uploading raw data. A third-party server coordinates the training of a global model by each user, and uses homomorphic encryption, differential privacy, etc. to encrypt the transmission of intermediate parameters. Compared with traditional diagnosis models, federated learning technology does not need to concentrate raw data, and there is no subsequent data transmission and public sharing, which can ensure data privacy. In actual industrial applications, this method still shows obvious shortcomings: 1) It ignores the case where only data exists on the client but lacks labels. There is no optimization for these customer groups. 2) It ignores the possibility that customers come from different clusters. Highly non-identically distributed data will seriously affect the performance of the model.

[0005] Therefore, the prior art needs to be further improved and perfected. SUMMARY

[0006] The application provides a dynamic weighted personalized federated meta-learning fault diagnosis method based on an unsupervised and grouping mechanism.

[0007] To achieve the above purpose, the application adopts the following technical solutions:

[0008] A dynamic weighted personalized federated meta-learning fault diagnosis method based on an unsupervised and grouping mechanism comprises the following steps:

[0009] S1, collect vibration signals of fault bearings under different devices (working conditions) as initial data sets of different clients;

[0010] S2, for the client with data but without labels, a pseudo-label algorithm based on a local outlier factor is used to label all sample signals with pseudo-labels. Then the server sends simulation data to the client, and the real label is obtained by calculating the Euclidean distance between the simulation data and the real data;

[0011] S3, for the possible non-independent and identically distributed data between clients, a personalized federated meta-learning algorithm based on a grouping mechanism is designed. Each data set will extract feature values such as peak value, average value and root mean square through feature engineering to construct a feature matrix, and then calculate the cluster center of the data set and upload it to the server. By calculating the distance between the cluster centers of different clients, the clients with similar features are placed in a group to train the global model;

[0012] S4, in the training process of each global model, a meta-learning strategy is introduced. The data set of each client is divided into a support set and a query set, and each client receives the global model and fine-tunes the new model on its own support set, and then tests the loss value on the query set. This loss value is used to update the global model;

[0013] S5, a dynamic weighted combination update parameter strategy is introduced. According to the designed formula, it is calculated how much the current client can benefit from other client models to effectively obtain the best weighted model combination of each client.

[0014] The above scheme further uses a pseudo-label algorithm based on a local outlier factor in step S2, and the overall process is as follows:

[0015] Assuming that there are multiple segments of unknown category signals on a certain client, a segment is first extracted as baseline data and is assigned a pseudo label, and then a segment is selected from outside the baseline data as comparison data. Then they are divided into the same number of subsamples. These subsamples form a comparison data set, and then the local outlier factor algorithm (LOF) is used for anomaly detection on the comparison data set. In theory, the abnormal sample proportion of the comparison data set composed of subsamples of two different category signals will be higher, so when the proportion is greater than the set threshold, it is determined that they are not of the same category, and different pseudo labels are assigned. When the proportion is less than the set threshold, it is determined that they are of the same category, and the same pseudo label is assigned. Then the next segment of signal is selected as comparison data. Through the above process cycle, the same pseudo label can be assigned to all signals of the same category.

[0016] The above scheme is further, the dynamic weighting combination updating parameter strategy process in the step S5 is as follows:

[0017] After each client fine-tunes its own model on the support set of local data, in order to further improve the performance of the local model, according to the formula, how much the current client can benefit from other client models is calculated, to effectively obtain the best weighted model combination of each client. The formula of the weight is as follows: ; wherein i is a local client, n is another client. represents the model parameter, represents the loss value obtained by testing the model in the query set of local data. If W n is negative, indicating that the model of the client n performs poorly on the data of the local client i , and its parameters will not be adopted. In addition, the denominator represents the difference between the model parameters of the local client and other clients, and the physical meaning is to avoid too large difference between the model parameters of different clients while focusing on the model performance, thereby improving the generalization ability and avoiding overfitting. Finally, the parameter update of the local client i is as follows: .

[0018] The application uses an unsupervised algorithm based on local outlier factor, so that unlabeled clients can also participate in federated learning smoothly. Meanwhile, in view of the problem that the fault data between clients exist highly dependent and identically distributed, a federated learning framework based on grouping mechanism is proposed, and meta-learning and dynamic weighting combination updating parameter strategy are introduced in the training of each global model; compared with the traditional federated learning method for fault diagnosis, not only the environment of various clients is considered, but also the performance of the model is improved, the bearing fault diagnosis accuracy is improved, the bearing health state can be diagnosed in time, and the economic loss is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flow chart of a dynamic weighted personalized federated meta-learning fault diagnosis method based on unsupervised and grouping mechanism;

[0020] Figure 2 is a bearing fault test bed used by the private data set of the present application;

[0021] Figure 3 is a schematic diagram of a pseudo-label algorithm based on local outlier factors designed by the present application;

[0022] Figure 4 is a grouping mechanism diagram in the present application;

[0023] Figure 5 is a schematic diagram of the meta-learning and dynamic weighted combined update parameter strategy in the present application;

[0024] Figure 6 is a comparison chart of the fault diagnosis method effect of the present application. DETAILED DESCRIPTION

[0025] The concept, specific structure and technical effects of the present application will be further described below in combination with the drawings to fully understand the purpose, features and effects of the present application.

[0026] Referring to Figures 1-6 , the present application relates to a dynamic weighted personalized federated meta-learning fault diagnosis method based on unsupervised and grouping mechanism, which comprises the following steps:

[0027] S1, collect the vibration signals of fault bearings under different equipment (working conditions) as the initial data set of different clients, wherein some clients have no labels.

[0028] S2, judge whether the data on the client has a label, for the client with data but without label, use the pseudo-label algorithm based on local outlier factors to label all sample signals. Then the server sends simulation data to this client, calculates the Euclidean distance between the simulation data and the real data, and obtains the real label.

[0029] S3, for the possible non-independent and identically distributed data between clients, design a personalized federated meta-learning algorithm based on grouping mechanism. Each data set will extract peak value, average value and root mean square value and other characteristic values through feature engineering to construct a feature matrix, then calculate the clustering center of the data set and upload it to the server. By calculating the distance between the clustering centers of different clients, the clients with similar features are placed in a group to train the global model.

[0030] S4, In the training process of each global model, a meta-learning strategy is introduced. The data set of each client is divided into a support set and a query set. After each client receives the global model, it fine-tunes the model on its own support set to obtain a new model, and then tests the loss value on the query set. This loss value is used to update the global model.

[0031] S5, a dynamic weighted combination update parameter strategy is introduced. According to the designed formula, it is calculated how much the current client can benefit from the models of other clients to effectively obtain the best weighted model combination of each client.

[0032] Further, the pseudo-label algorithm based on local outlier factor in step S2, assuming that there are multiple segments of unknown category signals on a certain client, first extract a segment as reference data and assign a pseudo-label, and then select a segment as comparison data from the reference data. Then divide them into the same number of sub-samples. These sub-samples form a comparison data set, and then use the local outlier factor algorithm (LOF) to perform anomaly detection on the comparison data set. In theory, the comparison data set composed of sub-samples of two different categories of signals will have a higher proportion of abnormal samples, so when the proportion is greater than the set threshold, it is determined that they are not of the same category, and are assigned different pseudo-labels. When the proportion is less than the set threshold, it is determined that they are of the same category, and are assigned the same pseudo-label. Then select the next segment of signal as comparison data. Through the above process cycle, the same pseudo-label can be assigned to all signals of the same category.

[0033] Then the server sends simulation data to this client, calculates the real category corresponding to the pseudo-label through the Euclidean distance to complete the virtual-real collaborative label correction, and can participate in the next step of federated learning training.

[0034] Further, the dynamic weighted combination update parameter strategy in step S5 has the following specific process:

[0035] After each client fine-tunes its own model on the support set of local data, in order to further improve the performance of the local model, according to the formula, it is calculated how much the current client can benefit from the models of other clients to effectively obtain the best weighted model combination of each client. The formula of the weight is as follows: ; wherein i is the local client, n is the other client. represents the model parameter, represents the loss value obtained by testing the model on the query set of local data. If W n is negative, indicating that the model of client n in the local client iThe data of the local client is poor, and the parameters thereof will not be adopted. In addition, the denominator represents the difference between the local client and other client model parameters, and the physical meaning is to avoid too large model parameter difference between different clients while paying attention to model performance, thereby improving the generalization ability and avoiding overfitting. Finally, the parameter update of the local client i is as follows: . Embodiments

[0036] The bearing open data set and the private data set of West Chester University are used to collect vibration signals of fault bearings under different working conditions, and the data sets of different clients are divided in a random proportion. Then, for the clients with data but without labels, a pseudo-label algorithm based on local outlier factors is used to label the same class data with the same pseudo-label, and the real label is obtained by calculating the Euclidean distance between the simulation data sent by the server and the real data. To solve the problem of non-independent and identically distributed data between clients, a personalized federated meta-learning algorithm based on grouping mechanism is designed, a feature matrix is constructed through feature engineering, and the clustering centers of different client data sets are calculated, so that the clients with similar features are grouped to train the global model. In the training process of the global model, the introduction of the meta-learning strategy enables the new model of each client to be fine-tuned on the support set and tested on the query set to update the global model. Finally, through the dynamic weighted combination update parameter strategy, the best weighted model combination of each client is calculated according to the designed formula, so as to improve the efficiency and accuracy of the model. And complete the fault diagnosis under cross-working condition.

[0037] Compared with the prior art, the beneficial effects of the technical scheme of the present application are:

[0038] (1) The present application proposes an unsupervised algorithm based on joint entropy, which solves the problem that unlabeled clients cannot participate in federated learning.

[0039] (2) The present application proposes a federated learning framework based on grouping mechanism, which greatly reduces the influence of the highly non-independent and identically distributed problem caused by fault data between clients.

[0040] (3) A personalized federated meta-learning algorithm based on dynamic weighted combination is proposed, which makes the model have stronger generalization ability.

[0041] Of course, the above drawings only describe the preferred embodiments of the present application, and those skilled in the art can make many improvements and changes to the above embodiments through logical analysis, reasoning or limited experiments without exceeding the concept and scope of the present application, and these improvements and changes should belong to the scope of protection claimed by the present application.

Claims

1. A Dynamically Weighted Personalized Federated Meta-Learning Based on Unsupervised Learning and Grouping Mechanisms The fault diagnosis method is characterized by comprising the following steps: S1. Collect vibration signals of faulty bearings under different equipment or operating conditions as the initial dataset for different clients; S2. For clients with data but no labels, a pseudo-labeling algorithm based on local outlier factors is used to label all sample signals with pseudo-labels. Then the server sends simulation data to this client and calculates the Euclidean distance between the simulation data and the real data to obtain the real labels. S3. To address the potential for non-independent and identically distributed data among customers, a personalized federated meta-learning algorithm based on a grouping mechanism is designed. Each dataset undergoes feature engineering to extract peak, average, and root mean square features to construct a feature matrix. Then, the cluster centers of the datasets are calculated and uploaded to the server. By calculating the distance between the cluster centers of different clients, clients with similar distances, i.e., similar features, are placed in a group to train the global model. S4. During the training process of each global model, a meta-learning strategy is introduced. Each client's dataset is divided into a support set and a query set. After each client receives the global model, it fine-tunes it on its own support set to obtain a new model, and then tests it on the query set to obtain a loss value. This loss value is used to update the global model. S5. A dynamic weighted combination parameter update strategy is introduced. Based on the designed formula, the extent to which the current client can benefit from other client models is calculated, thereby effectively obtaining the optimal weighted model combination for each client. The designed formula is as follows: ;in i For local clients, n For other clients, Represents model parameters, The loss value obtained by testing the model on the query set of local data, if W n A negative value indicates that the client... n The model on the local client i If the data shows poor performance, its parameters will not be adopted. Furthermore, the denominator represents the difference between the model parameters of the local client and other clients. The physical meaning is to avoid excessive differences in model parameters between different clients while focusing on model performance, thereby improving generalization ability and avoiding overfitting. Ultimately, the local client... i The parameters are updated as follows: .

2. The fault diagnosis method based on unsupervised and grouping mechanisms using dynamic weighted personalized federated meta-learning as described in claim 1, characterized in that, Step S2 uses a pseudo-label algorithm based on local outlier factors, and its overall process is as follows: Suppose a client has multiple signals of unknown category. First, extract one segment as the baseline data and assign it a pseudo-label. Then, select another segment from outside the baseline data as the comparison data. Divide both segments into the same number of subsamples. These subsamples form the comparison dataset. Then, use the Local Outlier Factor (LOF) algorithm to perform anomaly detection on the comparison dataset. Theoretically, the proportion of anomalous samples in the comparison dataset composed of subsamples of two different categories of signals will be higher. Therefore, when the proportion is greater than a set threshold, they are determined to be of different categories and assigned different pseudo-labels. When the proportion is less than the set threshold, they are determined to be of the same category and assigned the same pseudo-label. Then, select the next signal segment as the comparison data. By repeating the above process, all signals of the same category can be assigned the same pseudo-label. The server then sends simulation data to the client, calculates the Euclidean distance to obtain the real category corresponding to the pseudo-label, completes the label correction for virtual-real collaboration, and can then participate in the next step of federated learning training.

Images