Rotating machinery equipment fault diagnosis method based on double contrast federated semi-supervised learning

Abstract

The present invention discloses a method for fault diagnosis of rotating mechanical equipment based on double contrast federated semi-supervised learning, which belongs to the technical field of machine learning and includes the following steps: S1, the client initializes a local model; S2, the client uses a local data set for local training and then uploads it to a server; S3, the server aggregates the collected local prototypes and redistributes the updated global prototype to the client; S4, the client updates the global prototype and the local model; S5, repeats S2 to S4 until the model converges or reaches a predefined number of iterations to obtain a pre-trained model; S6, fine-tunes using a small amount of local labeled data; S7, inputs test data into the fine-tuned model to obtain a test result. The present invention uses a local data set for learning by multiple clients, and promotes joint learning between clients through lightweight prototypes, thereby achieving mutual knowledge sharing, preventing local model divergence, and improving the accuracy of equipment fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a rotating mechanical equipment fault diagnosis method based on double-contrast federated semi-supervised learning. Background Art

[0002] With the rapid development of modern industry, the complexity and intelligence level of large-scale industrial equipment have increased dramatically, and bearing fault diagnosis has become crucial. Intelligent fault diagnosis plays a vital role in modern industrial production. It is not only a key means to ensure the normal operation of equipment and improve production efficiency, but also an important support for promoting industrial technological progress and social and economic development.

[0003] In recent years, many deep learning-based intelligent fault diagnosis methods have been proposed, which utilize a large amount of vibration monitoring data collected from factories. However, most deep learning techniques adhere to the supervised learning paradigm, which requires obtaining a sufficient number of labeled samples to ensure stable convergence. However, obtaining high-quality labeled data in actual scenarios is a challenging task and usually requires a lot of human and material resources. In addition, industrial equipment is usually distributed in different geographical locations, forming data silos. In addition, the industrial field usually involves sensitive data, such as equipment operating status and production data, which brings the risk of privacy leakage when collecting and centralizing training data sets. In addition, changes in equipment conditions, working environment, and operating conditions in actual production environments may lead to differences in data distribution. Supervised deep learning methods are usually used to train models. Supervised models usually require a large amount of manually labeled data for training, but manual labels are often expensive and difficult to obtain. In the presence of label scarcity and data distribution differences, how to use local unlabeled data and features from different clients to improve training performance and improve the generalization ability of the model without leaking privacy remains a challenge.

[0004] Therefore, a rotating machinery equipment fault diagnosis method based on double contrast federated semi-supervised learning is proposed to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a rotating machinery equipment fault diagnosis method based on double-contrast federated semi-supervised learning, which uses a large amount of unlabeled data and a small amount of labeled data for learning through multiple clients, and promotes joint learning between clients through lightweight prototypes, thereby achieving mutual knowledge sharing, preventing local model divergence, and improving the accuracy of equipment fault diagnosis.

[0006] To achieve the above object, the present invention provides a rotating machinery equipment fault diagnosis method based on double contrast federated semi-supervised learning, comprising the following steps:

[0007] S1, the client initializes the local model;

[0008] S2, use the local data set to perform local training on the client, and upload the local prototype to the server after the training is completed;

[0009] S3, the server aggregates the collected local prototypes and redistributes the updated global prototypes to the client;

[0010] S4, the client updates the global prototype and local model;

[0011] S5, repeat S2 to S4 until the model converges or reaches a predefined number of iterations to obtain a pre-trained model;

[0012] S6: The client uses the pre-trained model obtained in S5 to fine-tune it using a small amount of local labeled data.

[0013] S7. Input the test data into the fine-tuned model to obtain the test results.

[0014] Preferably, in S2, the local data includes a large amount of unlabeled data and a small amount of labeled data, wherein the amount of unlabeled data is significantly greater than that of labeled data; in addition, the data between different clients are not independent and identically distributed;

[0015] The server refers to a coordinator that can perform basic operations, establish and initialize global prototypes, and receive local prototypes uploaded by each participant.

[0016] Preferably, in S2, the specific process of local training is: first, data enhancement is performed on unlabeled data and labeled data respectively, and data enhancement is divided into strong enhancement A(·) and weak enhancement α(·). Data enhancement is a method of expanding data by performing time domain transformation on data. The time domain transformation includes Gaussian noise, sampling, subsequence permutation, etc., wherein strong enhancement and weak enhancement are random selection or random combination of the above methods;

[0017] Then, the supervised loss function of the labeled data and the unsupervised loss function of the unlabeled data are calculated respectively. The specific steps are as follows: weakly enhance the labeled data into the model Get its prediction, And calculate the supervised loss:

[0018]

[0019] Where E is the number of labeled samples, H(p,q) is the cross entropy loss, yes The true label of

[0020] At the same time, for unlabeled data, it is strongly enhanced A(u i ) and weak enhancement α(ui ) input model, and record the weakly enhanced output as Will As pseudo labels, calculate the unlabeled loss function, that is, calculate the prediction of pseudo labels and strong enhancement The cross entropy between:

[0021]

[0022] Then, for the strong and weak enhancements of unlabeled data, the features extracted by the model for strong enhancement are recorded as c s , the features extracted by the model for weak enhancement are recorded as c w , calculate the local contrast loss:

[0023]

[0024] in, represents the cosine similarity between x and y, B is the batch size, τ i and τ j is the temperature coefficient, is an indicator function, which takes the value 0 when i=j and takes the value 1 when i≠j;

[0025] For the local prototype and the global prototype, the global contrast loss is calculated, and the expression is:

[0026]

[0027] Where C is the number of categories, is an indicator function, which takes the value 0 when k=j and takes the value 1 when k≠j;

[0028] Finally, the total loss function in the training process is obtained, which is expressed as:

[0029] L total =L s +η(t)L u +L LC +ι(t)L GC

[0030] Among them, L total is the loss function in, L s is the supervised loss function, L u is the unsupervised loss function, L LC is the local contrast loss, L GC is the global contrast loss; η(t) is the weight coefficient of the unsupervised loss function, and ι(t) is the weight coefficient of the global contrast loss.

[0031] Preferably, the expression of η(t) is:

[0032]

[0033] Among them, T 1 =0.3×T,T 2 =0.7×T, T is the number of communication rounds;

[0034] ι(t)=E / B

[0035] Among them, E is the number of labeled samples and B is the batch size.

[0036] Preferably, in S2, the local prototype is the average of the features of each category, and the local prototype of the jth category of the kth user is The calculation formula is as follows:

[0037]

[0038] Among them, D k,j Indicates that the k-th user contains the sub-dataset of the j-th category, f k (θ k ; x) represents the encoder part of the k-th user model, θ k are the parameters of the encoder.

[0039] Preferably, in S3, aggregating the collected local prototypes means that given a category j, the server collects prototypes from clients with the jth category and performs weighted aggregation to generate a global prototype of the jth category. The calculation formula is as follows:

[0040]

[0041] Among them, N j represents the total number of samples of the jth category of all users, represents the set of clients that have samples of the jth type;

[0042] The global prototype of the tth round is a collection of C categories of global prototypes, expressed as:

[0043]

[0044] Preferably, in S3, the updated global prototype refers to the global prototype after exponential moving average update, and the calculation formula is as follows:

[0045]

[0046] Among them, P t is the global prototype after the tth round of update, and κ is the weight coefficient.

[0047] Therefore, the present invention is based on a rotating machinery equipment fault diagnosis method based on double-contrast federated semi-supervised learning. By using semi-supervised learning, representation learning, and prototype-based federated learning, effective representations can be learned on the sparsely labeled data of clients with different distributions to train the model, and performance similar to that of supervised learning can be obtained, and even the performance of supervised learning can be surpassed. This ensures the accuracy of the model while reducing the amount of data annotation, and can effectively solve the problems of different data distributions, difficulty in obtaining labels, no labels, and insufficient labels; the model accuracy and generalization ability are improved, thereby improving the accuracy of equipment fault diagnosis.

[0048] The method proposed in this invention performs well in environments with insufficient labeled data and significant differences in data distribution. It effectively extracts key representations of samples through comparative learning and realizes collaborative learning among different factories through prototype knowledge sharing. This method not only improves the accuracy and generalization ability of model training, but also significantly reduces the dependence on manual expertise, enabling it to effectively cope with complex fault diagnosis tasks. This technology has extremely broad application prospects in the field of intelligent fault diagnosis of large rotating machinery.

[0049] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the process of the present invention;

[0051] Figure 2 The figure is a schematic diagram of the flow chart of motor bearing fault diagnosis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0054] The present invention provides a rotating mechanical equipment fault diagnosis method based on double contrast federated semi-supervised learning, such as Figure 1 As shown, the following steps are included:

[0055] S1, the client initializes the local model;

[0056] S2. Perform local training on the client using a local data set. The local data includes a large amount of unlabeled data and a small amount of labeled data. The amount of unlabeled data is significantly greater than that of labeled data. In addition, the data between different clients are not independent and identically distributed. After completing the training, upload the local prototype to the server.

[0057] The specific process of local training is as follows: first, data enhancement is performed on unlabeled data and labeled data respectively. Data enhancement is divided into strong enhancement A(·) and weak enhancement α(·). Data enhancement is a method of expanding data by performing time domain transformation on data. Time domain transformation includes Gaussian noise, sampling, subsequence permutation, etc. Among them, strong enhancement and weak enhancement are random selection or random combination of the above methods;

[0058] Secondly, the weakly enhanced input of labeled data into the model Get its prediction, And calculate the supervised loss:

[0059]

[0060] Where E is the number of labeled samples, H(p,q) is the cross entropy loss, yes The true label of

[0061] At the same time, for unlabeled data, it is strongly enhanced A(u i ) and weak enhancement α(u i ) input model, and record the weakly enhanced output as Will As pseudo labels, calculate the unlabeled loss function, that is, calculate the prediction of pseudo labels and strong enhancement The cross entropy between:

[0062]

[0063] Then, for the strong and weak enhancements of unlabeled data, the features extracted by the model for strong enhancement are recorded as c s , the features extracted by the model for weak enhancement are recorded as c w , calculate the local contrast loss:

[0064]

[0065] in, represents the cosine similarity between x and y, B is the batch size, τ i and τ j is the temperature coefficient, is an indicator function, which takes the value 0 when i=j and takes the value 1 when i≠j;

[0066] For local prototypes and the global prototype Calculate the global contrast loss:

[0067]

[0068] Where C is the number of categories, is an indicator function, which takes the value 0 when k=j and takes the value 1 when k≠j;

[0069] The local prototype is the average of the features of each category. The local prototype of the jth category of the kth user is The calculation formula is as follows:

[0070]

[0071] Among them, D k,j Indicates that the k-th user contains the sub-dataset of the j-th category, f k (θ k ; x) represents the encoder part of the k-th user model, θ k are the parameters of the encoder.

[0072] Finally, the total loss function in the training process is obtained, which is expressed as:

[0073] L total =L s +η(t)L u +L LC +ι(t)L GC

[0074] Among them, L total is the loss function in, L s is the supervised loss function, L u is the unsupervised loss function, L LC is the local contrast loss, L GC is the global contrast loss; η(t) is the weight coefficient of the unsupervised loss function, and the expression of η(t) is:

[0075]

[0076] Among them, T 1 =0.3×T,T 2 =0.7×T, T is the number of communication rounds;

[0077] ι(t) is the weight coefficient of the global contrast loss, expressed as:

[0078] ι(t)=E / B

[0079] Among them, E is the number of labeled samples and B is the batch size.

[0080] S3, the server aggregates the collected local prototypes and redistributes the updated global prototypes to the client;

[0081] Aggregating the collected local prototypes means that given a category j, the server collects prototypes from clients with the jth category and performs weighted aggregation to generate a global prototype of the jth category. The calculation formula is as follows:

[0082]

[0083] Among them, N j represents the total number of samples of the jth category of all users, represents the set of clients that have samples of the jth type;

[0084] The global prototype of the tth round is a collection of C categories of global prototypes, expressed as:

[0085]

[0086] The updated global prototype refers to the global prototype after exponential sliding average update, and the calculation formula is as follows:

[0087]

[0088] Among them, P t is the global prototype after the tth round of update, and κ is the weight coefficient.

[0089] S4, update the global prototype and local model on the client;

[0090] S5, repeat S2 to S4 until the model converges or reaches a predefined number of iterations to obtain a pre-trained model;

[0091] S6: The client uses the pre-trained model obtained in S5 to fine-tune it using a small amount of local labeled data.

[0092] S7. Input the test data into the fine-tuned model to obtain the test results.

[0093] Example

[0094] This embodiment provides an intelligent fault diagnosis method for motor bearings:

[0095] With the rapid development of modern industry, large-scale industrial equipment has become increasingly complex and intelligent, which makes bearing fault diagnosis particularly critical. Accurate bearing fault diagnosis plays a vital role in ensuring stable operation of equipment, improving production efficiency, and promoting industrial development and social and economic progress.

[0096] like Figure 2As shown, this embodiment demonstrates a framework for intelligent fault diagnosis of local industrial data of each factory. The data characteristics of each factory include insufficient labels and differences in data distribution between different factories. During the training phase, each factory independently trains a local model using its local data, and then uploads the trained local prototype to the central server. The server aggregates all local prototypes, updates the global prototype, and redistributes it to the clients of each project for further training and updating of local models. This process is repeated until the model converges or reaches a preset number of iterations, thereby providing a personalized pre-trained model for each factory.

[0097] In the application phase, each factory first fine-tunes the pre-trained model using a small amount of labeled local data. Subsequently, the test data is input into the fine-tuned model to obtain the fault diagnosis results. This fault diagnosis process is fast and easy, and its accuracy and reliability do not rely on human knowledge and experience, thus ensuring the efficiency and objectivity of the diagnosis process.

[0098] Therefore, the present invention is based on a rotating machinery equipment fault diagnosis method based on double-contrast federated semi-supervised learning. Multiple clients use a large amount of unlabeled data and a small amount of labeled data for learning, and promotes joint learning between clients through lightweight prototypes, thereby achieving mutual knowledge sharing, preventing local model divergence, and improving the accuracy of equipment fault diagnosis.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for fault diagnosis of rotating machinery equipment based on double contrast federated semi-supervised learning, characterized in that: The following steps are involved: S1, the client initializes the local model; S2, use the local data set to perform local training on the client, and upload the local prototype to the server after training; the local data includes a large amount of unlabeled data and a small amount of labeled data; The specific process of local training is as follows: first, data enhancement is performed on unlabeled data and labeled data respectively, and then the supervised loss function is calculated for labeled data. For unlabeled data, the unsupervised loss function is calculated based on the generated pseudo labels. Then, local contrast loss is calculated for unlabeled data, and global contrast loss is calculated for local prototypes and global prototypes. Finally, the total loss function in the training process is obtained, which is expressed as: L total =L s +η(t)L u +L LC +ι(t)L GC Among them, L total is the loss function in, L s is the supervised loss function, L u is the unsupervised loss function, L LC is the local contrast loss, L GC is the global contrast loss; η(t) is the weight coefficient of the unsupervised loss function, ι(t) is the weight coefficient of the global contrast loss; The expression of η(t) is: Where, T1 = 0.3 × T, T2 = 0.7 × T, T is the number of communication rounds; ι(t)=E / B Among them, E is the number of labeled samples and B is the batch size; S3, the server aggregates the collected local prototypes and redistributes the updated global prototypes to the client; S4, the client updates the global prototype and local model; S5, repeat S2 to S4 until the model converges or reaches a predefined number of iterations to obtain a pre-trained model; S6: The client uses the pre-trained model obtained in S5 to fine-tune it using a small amount of local labeled data. S7. Input the test data into the fine-tuned model to obtain the test results.

2. The method for fault diagnosis of rotating machinery equipment based on double contrast federated semi-supervised learning according to claim 1 is characterized in that: In S2, the local prototype is the average of the features of each category, and the local prototype of the jth category of the kth user is The calculation formula is as follows: Among them, D k，j Indicates that the k-th user contains the sub-dataset of the j-th category, f k (θ k ; x) represents the encoder part of the k-th user model, θ k are the parameters of the encoder.

3. The method for fault diagnosis of rotating machinery equipment based on double contrast federated semi-supervised learning according to claim 2 is characterized in that: In S3, aggregating the collected local prototypes means that given a category j, the server collects prototypes from clients with the ,th class, and performs weighted aggregation to generate a ,th class global prototype, the ,th class global prototype The calculation formula is as follows: Among them, N j represents the total number of samples of the jth category of all users, represents the set of clients that have samples of the jth type; The global prototype of the tth round is a collection of C categories of global prototypes, expressed as:

4. The method for fault diagnosis of rotating machinery equipment based on double contrast federated semi-supervised learning according to claim 3 is characterized in that: In S3, the updated global prototype refers to the global prototype after exponential moving average update, and the calculation formula is as follows: Among them, P t is the global prototype after the tth round of update, and κ is the weight coefficient.

Images