CN116910652A - An equipment fault diagnosis method based on federated self-supervised learning - Google Patents

Abstract

The invention discloses a device fault diagnosis method based on federal self-supervision learning. Comprising the following steps: firstly, initializing the weight of a feature extractor by a server and transmitting the weight to each client; then, each client acquires signals generated when the local equipment works by using a sensor and records the signals as local vibration data, so as to obtain a non-tag data set and a tag data set; then, each client trains the local feature extractor under the federal self-supervision learning framework respectively, and further obtains the trained local feature extractor; respectively training classifiers by each client under a supervision and learning framework to obtain corresponding client classifiers, and connecting the feature extractor and the client classifiers to form a fault diagnosis model; and finally, performing equipment diagnosis by using a fault diagnosis model of the client. The invention solves the problems that the fault data set of the rotating equipment is smaller and dispersed, and the high-precision diagnosis model is difficult to train due to the lack of labels.

Description

An equipment fault diagnosis method based on federated self-supervised learning

Technical field

The invention belongs to the field of rotating equipment fault diagnosis, relates to an equipment fault diagnosis method in the fields of machine learning, deep learning and time series classification, and specifically relates to an equipment fault diagnosis method based on federated self-supervised learning.

Background technique

Rotating equipment is widely used in modern industry and is becoming more and more complex and sophisticated, such as aircraft engines and gas turbines. When rotating equipment has faults such as bearing damage or blade breakage, it may cause serious accidents and cause huge economic losses. Therefore, correctly identifying the status of equipment in operation and promptly intervening when early signs of failure occur are of great significance to improving production efficiency and reducing disaster losses.

Most of the current rotating equipment fault diagnosis methods are data-driven. Through a large amount of data, the mapping relationship between vibration signals and equipment status is mined. Commonly used methods include support vector machines, decision trees, and neural networks. Deep neural network is one of the current mainstream research directions because of its ability to automatically extract features. For example, the method disclosed in the patent "A Machine Learning-Based Fault Diagnosis Method for Rotating Equipment in Cement Production" uses a one-dimensional convolutional neural network and a fully connected neural network to extract vibration features, and then uses integrated learning to obtain diagnostic results from multiple classifiers. . The method disclosed in the patent "A rotating equipment fault diagnosis method, system and readable storage medium based on a deep residual network" uses a deep residual network to extract fault features from vibration signals.

Although existing methods have achieved a very high level of diagnostic accuracy, they use large-size, fully-labeled data sets to train models, which is limited in many situations. Lack of annotation of data is the most common limitation. Vibration signals need to be annotated by experts with domain knowledge, which is very costly. In practical applications, only a small part of the data is annotated. Data security also needs to be considered. Different customers use equipment under different working conditions. To ensure model robustness, data from all customers should be collected as much as possible. However, customers may be unwilling to share their data to train better models out of interest considerations or concerns about the risk of data leakage.

Contents of the invention

In order to solve the problems and needs existing in the background technology, the present invention proposes an equipment fault diagnosis method based on federated self-supervised learning, which can utilize multiple small data sets lacking labels that are scattered and cannot be shared to train a fault diagnosis model and use it online diagnosis. Use self-supervised learning methods to learn effective fault feature representations from a large amount of unlabeled data, and use federated learning methods to train fault diagnosis models with global knowledge from multiple clients without requiring clients to upload local data. Train efficient fault diagnosis models in scenarios where the data set is small, scattered and lacks labels.

The specific technical solution of the present invention includes several steps:

S1: The server initializes the weight of the feature extractor and sends the weight to each client. Each client uses this weight as the initial weight of the local feature extractor;

S2: Each client uses sensors to collect signals generated by local equipment during operation and records them as local vibration data. Then, after preprocessing the local vibration data, obtain unlabeled data sets and labeled data sets;

S3: Under the federated self-supervised learning framework, each client uses an unlabeled data set to train a local feature extractor, and then obtains a trained local feature extractor;

S4: Each client uses the labeled data set to train a classifier under the supervised learning framework to obtain the corresponding client classifier. In each client, a fault diagnosis model is formed by connecting the current feature extractor and the client classifier;

S5: After preprocessing, the sensor data of the device to be diagnosed is connected to the fault diagnosis model of the corresponding client to obtain the corresponding device diagnosis results.

In the S1, the feature extractor uses a convolutional neural network with residual connections.

In the S2, each client performs the following steps:

S21: Use the accelerometer to collect the signals generated by the local equipment during operation and record them as local vibration data;

S22: Randomly select a preset proportion of local vibration data and label the selected local vibration data according to the real state of the device to obtain an initial labeled data set. The unselected local vibration data is recorded as an initial unlabeled data set;

S23: Use a sliding window to divide all signals of the initial labeled data set and unlabeled data set into multiple segments, and obtain the segmented labeled data set and unlabeled data set respectively;

S24: Use the max-min method to numerically scale the segmented labeled data set and unlabeled data set, respectively, to obtain the final unlabeled data set and labeled data set.

The S3 is specifically:

S31: In each round of training, each client uses an unlabeled data set to train a local feature extractor under a self-supervised learning framework. The local feature extractor weights of each client after the current round of training are obtained and uploaded to the server;

S32: After the server aggregates the local feature extractor weights of all clients, it obtains the global feature extractor weight and delivers the weight to the local feature extractor of each client;

S33: Repeat S31-S32 to update the global feature extractor weights for multiple rounds until the preset round is reached. The final global feature extractor weights are sent to the local feature extractors of each client, so that each client obtains training. Good local feature extractor.

In the S31, each client performs the following steps:

S311: Use two different data enhancement methods to perform data enhancement on each piece of unlabeled data in the unlabeled data set, and obtain the corresponding enhanced sample pairs;

S312: Under the self-supervised learning framework, use the enhanced samples corresponding to the unlabeled data set to perform a round of local feature extractor training on the data set, obtain the weight of the local feature extractor after the current round of training and upload it to the server.

In the S311, the first data enhancement sample It is obtained by first adding Gaussian noise to each piece of unlabeled data, and then scaling the noise-added data; the second data enhancement sample/> It is obtained by first smoothly warping the interval of time steps of each piece of unlabeled data and then applying noise.

During the training process, the local feature extractor outputs the first feature matrix and the first characteristic matrix/> Use the gradient descent algorithm to optimize the weight of the feature extractor. The loss function includes the first loss function and the second loss function. The calculation formula of the first loss function loss1 is as follows:

Among them, N is the batch size, α and β are the first and second hyperparameters respectively, is the indicator function; i and j represent the first and second index of the sample in the batch respectively; l _c is the context contrast loss function, l _t is the time contrast loss function/> Represents the first clipping feature matrix, /> represents the second cropping feature matrix, s1 and s2 represent the first and second starting positions respectively, e1 and e2 represent the first and second ending positions respectively; T represents the length of the time dimension of the feature matrix;/> Represents the second clipping feature matrix/> Characteristics of sample i at time step t′,/> Represents the first clipping feature matrix/> Characteristics of sample i at time step t,/> Represents the first clipping feature matrix/> Sample i in ,/> Represents the second clipping feature matrix/> Sample i in ,/> Represents the second clipping feature matrix/> Sample j in

The calculation formula of the second loss function loss2 is as follows:

in, Respectively represent the first feature matrix extracted by the local feature extractor of the k-th client in the r-th round/> and the first characteristic matrix/> Respectively represent the first feature matrix extracted by the local feature extractor of the p-th round of the k-th client/> and the first characteristic matrix/> Represents the features extracted by each client using the global feature extractor weights received from the server at round r of training.

In S32, the server uses the weighted average method to aggregate the local feature extractor weights of all clients. The calculation formula is as follows:

Among them, k is the client index, |D|, | ^Dk | represent the data volume of the global client and the data volume of the k-th client respectively, θ ^G and θ ^k represent the weight of the global feature extractor and the k-th client respectively. Local feature extractor weights at the end.

In the S4, each client performs the following steps:

S51: Input the label data set into the local feature extractor with updated weights to obtain the feature matrix data set;

S52: Use the feature matrix data set as the input of the support vector machine and train the support vector machine to obtain a client classifier.

In S5, a sliding window is used to perform signal segmentation on the sensor data of the device to be diagnosed, and a maximum-minimum method is used for normalization processing before being input into the fault diagnosis model of the corresponding client.

In the method of the present invention, federated learning is used to train a feature extractor that is effective for all clients from scattered client data; during the training process, self-supervised learning is used to learn useful knowledge from a large amount of unlabeled data; supervised learning is used Improving the final performance of classifiers from small amounts of labeled data.

Compared with the existing technology, the beneficial effects of the present invention are:

1. Model training is all performed locally on the client. Customer data does not need to be uploaded to the server, and customers do not need to worry about data leakage.

2. The method of the present invention can learn knowledge from a large amount of unlabeled data, and can exert good performance in the current situation where labels are generally lacking in equipment fault diagnosis applications.

3. The present invention uses comparative self-supervised learning to train a robust model from a small unlabeled data set; it uses federated learning to aggregate fault feature extractors with global knowledge from multiple clients, and avoids client local data sharing, thereby It solves the problem that the rotating equipment fault data set is small, scattered and lacks labels, making it difficult to train a high-precision diagnosis model.

Description of the drawings

Figure 1 is a schematic flow chart of specific steps of the present invention.

Figure 2 is a schematic diagram of the training and use of the model in the present invention.

Figure 3 is a schematic diagram of the experimental bench in the embodiment of the present invention.

Figure 4 shows the data distribution of three clients in the embodiment of the present invention.

Figure 5 is a confusion matrix of the first client fault diagnosis result in the embodiment of the present invention.

Figure 6 is a confusion matrix of the second client fault diagnosis result in the embodiment of the present invention.

Figure 7 is a confusion matrix of the third client fault diagnosis result in the embodiment of the present invention.

Detailed ways

The present invention will be further described below using the Case Western Reserve University Bearing Failure Dataset (CWRU) as the data of a specific embodiment.

The testbed of the CWRU data set consists of a motor, a torque sensor and a power tester. The rotating shaft of the motor is supported by the bearing to be tested. An accelerometer is installed on the motor and the vibration signal is sampled for one second at a frequency of 12KHZ. The experimental bench diagram of this data set is shown in Figure 3. There are three types of bearing failures, namely inner ring damage, outer ring damage and roller damage. Each fault type can also be broken down into three severity levels. Therefore, including the normal state, there are ten states in total for bearing data.

Figures 1 and 2 show the process of the present invention, combined with the implementation of the CWRU data set, which specifically includes the following steps:

S1: The server initializes the weight of the feature extractor and sends the weight to each client. Each client uses this weight as the initial weight of the local feature extractor;

In S1, the feature extractor uses a convolutional neural network (ResNet) with residual connections, which is a widely used deep learning model. The feature extractor structure of the server and client is consistent. In order to adapt to time series data, the convolution kernel of Resnet is restricted to slide only along the time dimension.

S2: Each client uses sensors to collect signals generated by local equipment during operation and records them as local vibration data. Then, after preprocessing the local vibration data, obtain unlabeled data sets and labeled data sets;

In S2, each client performs the following steps:

S21: Use the accelerometer to collect the signals generated by the local equipment during operation and record them as local vibration data; specifically, as shown in Figure 3, install the accelerometer near the motor bearing. Run the motor, and at a later point start the accelerometer to sample the signal for ten seconds and record the bearing status.

S22: Randomly select local vibration data with a preset ratio and label the selected local vibration data according to the real status of the device reflected by the data. Divide the selected local vibration data into an initial labeled data set and an initial test set according to a ratio of 1:2. Unselected local vibration data The vibration data is recorded as the initial unlabeled data set; in the specific implementation, the preset proportion is set to 30%.

S23: Use a sliding window with a length of 1024 and a sliding step of 1024 to segment all the signals of the initial labeled data set, the initial unlabeled data set and the initial test set into non-overlapping segments, and obtain the segmented data respectively. There are labeled data sets, unlabeled data sets and test sets; the CWRU data set contains vibration data of 161 bearings. This embodiment uses three clients as an example to illustrate. Therefore, the vibration data of 161 bearings are divided and divided according to non-independent synchronization. The distribution method is divided into local data sets of these three clients. The data distribution is shown in Figure 4. It can be seen that the data distribution of the three clients is quite different. The data volume ratio of each client's local labeled data set, unlabeled data set and test set is approximately 1:7:2.

S24: Use the max-min method to numerically scale the segmented labeled data set, unlabeled data set and test set to the [0,1] interval to obtain the final unlabeled data set and labeled data set. and test set;

S3: Under the federated self-supervised learning framework, each client uses an unlabeled data set to train a local feature extractor, and then obtains a trained local feature extractor;

S3 specifically is:

S31: In each round of training, each client uses an unlabeled data set to train a local feature extractor under a self-supervised learning framework. The local feature extractor weights of each client after the current round of training are obtained and uploaded to the server;

In S31, each client performs the following steps:

S311: Use two different data enhancement methods to perform data enhancement on each piece of unlabeled data in the unlabeled data set, and obtain the corresponding enhanced sample pairs;

In S311, the first data enhancement sample It is obtained by first adding Gaussian noise to each piece of unlabeled data, and then scaling the noise-added data; the second data enhancement sample/> It is obtained by first smoothly warping the interval of time steps of each piece of unlabeled data and then applying noise. Express a batch of unlabeled data sets as: x _train = {x ¹ , x ² ,..., x ^N }, where x∈R ^L and N is the batch size. In this embodiment, a batch of data is sampled from a standard Gaussian distribution using the same noise and scaling factors.

S312: Under the self-supervised learning framework, use the enhanced samples corresponding to the unlabeled data set to perform a round of local feature extractor training on the data set, obtain the weight of the local feature extractor after the current round of training and upload it to the server. In this embodiment, ResNet is used to extract feature representations of enhanced samples. The convolution kernel size of Resnet is fixed to [3,1] and is restricted to sliding only along the time dimension. The output dimension of Resnet, that is, the length of the feature representation, is set to 64.

During the training process, the local feature extractor f outputs the first feature matrix and the first characteristic matrix/> T represents the length of the time dimension of the first feature matrix, that is, the number of time steps. The gradient descent algorithm is used to optimize the weight of the feature extractor. In this embodiment, the Adam optimizer is used to perform the gradient descent algorithm, and the learning rate is set to 3e ^-4 . The loss function includes the first loss function and the second loss function. The calculation formula of the first loss function loss1 is as follows:

Among them, N is the batch size, α and β are the first and second hyperparameters respectively, which are used to adjust the contribution weight of l _c and l _t . is an indicator function, the value is 1 when the conditions in parentheses are true, otherwise it is 0; i and j respectively represent the first and second index of the sample in the batch; l _c is the context contrast loss function, the purpose of this loss function is to narrow the features The feature distance between the enhanced sample pairs output by the extractor also increases the feature distance between the enhanced sample and all other samples; l _t is the time contrast loss function, which further constrains the feature extractor from the similarity of time steps. output. Specifically, try to minimize the feature distance between the enhanced sample pairs at the same time step position, try to expand the feature distance between the enhanced sample pairs at different time step positions, and try to expand the feature distance at different time steps of each sample itself. l _c and l _t are contrastive loss terms designed based on the idea of NT-Xent loss function. They use the enhanced context dependence and time dependence between samples in batch data to construct supervision information respectively, thereby helping the feature extractor benefit from unlabeled data. . For the first characteristic matrix/> and the second characteristic matrix/> Apply random cropping to split into shorter paragraphs/> Represents the first cropping feature matrix, which is the pair of first feature matrices from the first starting position s1 to the first ending position e1/> The matrix after clipping,/> Represents the second cropping feature matrix, which is the pair of second feature matrices from the first starting position s2 to the first ending position e2/> The matrix after clipping,/> with/> The [s2:e1] part overlaps, s1 and s2 represent the first and second starting positions respectively, e1 and e2 represent the first and second ending positions respectively; T represents the length of the time dimension of the feature matrix, that is, the time step Quantity;/> Represents the second clipping feature matrix/> Characteristics of sample i at time step t′,/> Represents the first clipping feature matrix/> Characteristics of sample i at time step t,/> Represents the first clipping feature matrix/> Sample i in ,/> Represents the second clipping feature matrix/> Sample i in ,/> Represents the second clipping feature matrix/> Sample j in

The second loss function loss2 is used to constrain the distance between the feature extractor and the feature extractor of the previous training round. The calculation formula of the second loss function loss2 is as follows:

in, Respectively represent the first feature matrix extracted by the local feature extractor of the k-th client in the r-th round/> and the first characteristic matrix/> Respectively represent the first feature matrix extracted by the local feature extractor of the p-th round of the k-th client/> and the first characteristic matrix/> Represents the features extracted by each client using the global feature extractor weights received from the server at the start of training in round r.

In this embodiment, the length of the feature representation that is retained by cropping is 32, the data cropping starting position of a batch is the same, and it is sampled from the uniform distribution U[0,32].

S32: After the server aggregates the local feature extractor weights of all clients, it obtains the global feature extractor weight and delivers the weight to the local feature extractor of each client as the initial weight of the local feature extractor of each client for the next training round. ;

In S32, the server uses the weighted average method to aggregate the local feature extractor weights of all clients. The calculation formula is as follows:

Among them, k is the client index, |D|, | ^Dk | represent the data volume of the global client and the data volume of the k-th client respectively, θ ^G and θ ^k represent the weight of the global feature extractor and the k-th client respectively. Local feature extractor weights at the end.

S33: Repeat S31-S32 to update the global feature extractor weights for multiple rounds until the preset round is reached. The final global feature extractor weights are sent to the local feature extractors of each client, so that each client obtains training. Good local feature extractor.

In this embodiment, after all three clients have completed feature extractor training and uploaded model weights, the server aggregates these model weights to form new model weights and sends them to the clients. The number of training times is set to 40.

S4: Each client uses the labeled data set to train a classifier under the supervised learning framework to obtain the corresponding client classifier. In each client, a fault diagnosis model is formed by connecting the current feature extractor and the client classifier;

In S4, each client performs the following steps:

S41: Input the label data set into the local feature extractor with updated weights to obtain the feature matrix data set;

S42: Use the feature matrix data set as the input of the support vector machine and train the support vector machine to obtain a client classifier.

S5: After preprocessing, the sensor data of the device to be diagnosed is connected to the fault diagnosis model of the corresponding client to obtain the corresponding device diagnosis results.

In this embodiment, the test set obtained in step S24 is actually the preprocessed data to be diagnosed. Use the fault diagnosis model to classify the test set data and obtain the equipment diagnosis results.

The device diagnosis results were compared with the actual results, and accuracy (Acc) and macro-average F1 score (MF1) were selected as evaluation indicators to measure model performance. In this embodiment, the above index results are shown in Table 1.

Table 1 is the model performance evaluation table

The evaluation results show that the method of the present invention can successfully diagnose the equipment and achieve high diagnostic accuracy, and the method is feasible and effective. Figure 5, Figure 6, and Figure 7 show the confusion matrices of the three client diagnosis results. It can be seen that although there are large differences in the data distribution of each client, the method of the present invention can achieve extremely high diagnostic accuracy.

The above embodiment is an implementation of the present invention on the bearing fault data set of Case Western Reserve University. However, the specific implementation of the fault diagnosis method of the present invention is not limited to bearings. Any equipment operation data collected through sensors can be collected in accordance with the principles and ideas of the present invention. Similar solutions for equipment fault diagnosis should be regarded as the scope of protection of the patent of this invention.

Claims (10)

1. An equipment fault diagnosis method based on federated self-supervised learning, which is characterized by including the following steps: S1: The server initializes the weight of the feature extractor and sends the weight to each client. Each client uses this weight as the initial weight of the local feature extractor; S2: Each client uses sensors to collect signals generated by local equipment during operation and records them as local vibration data. Then, after preprocessing the local vibration data, obtain unlabeled data sets and labeled data sets; S3: Under the federated self-supervised learning framework, each client uses an unlabeled data set to train a local feature extractor, and then obtains a trained local feature extractor; S4: Each client uses the labeled data set to train a classifier under the supervised learning framework to obtain the corresponding client classifier. In each client, a fault diagnosis model is formed by connecting the current feature extractor and the client classifier; S5: After preprocessing, the sensor data of the device to be diagnosed is connected to the fault diagnosis model of the corresponding client to obtain the corresponding device diagnosis results. 2. An equipment fault diagnosis method based on federated self-supervised learning according to claim 1, characterized in that in S1, the feature extractor adopts a convolutional neural network with residual connections. 3. A device fault diagnosis method based on federated self-supervised learning according to claim 1, characterized in that, in the S2, each client performs the following steps: S21: Use the accelerometer to collect the signals generated by the local equipment during operation and record them as local vibration data; S22: Randomly select a preset proportion of local vibration data and label the selected local vibration data according to the real state of the device to obtain an initial labeled data set. The unselected local vibration data is recorded as an initial unlabeled data set; S23: Use a sliding window to divide all signals of the initial labeled data set and unlabeled data set into multiple segments, and obtain the segmented labeled data set and unlabeled data set respectively; S24: Use the max-min method to numerically scale the segmented labeled data set and unlabeled data set, respectively, to obtain the final unlabeled data set and labeled data set. 4. An equipment fault diagnosis method based on federated self-supervised learning according to claim 1, characterized in that the S3 is specifically: S31: In each round of training, each client uses an unlabeled data set to train a local feature extractor under a self-supervised learning framework. The local feature extractor weights of each client after the current round of training are obtained and uploaded to the server; S32: After the server aggregates the local feature extractor weights of all clients, it obtains the global feature extractor weight and delivers the weight to the local feature extractor of each client; S33: Repeat S31-S32 to update the global feature extractor weights for multiple rounds until the preset round is reached. The final global feature extractor weights are sent to the local feature extractors of each client, so that each client obtains training. Good local feature extractor. 5. A device fault diagnosis method based on federated self-supervised learning according to claim 4, characterized in that, in the S31, each client performs the following steps: S311: Use two different data enhancement methods to perform data enhancement on each piece of unlabeled data in the unlabeled data set, and obtain the corresponding enhanced sample pairs; S312: Under the self-supervised learning framework, use the enhanced samples corresponding to the unlabeled data set to perform a round of local feature extractor training on the data set, obtain the weight of the local feature extractor after the current round of training and upload it to the server. 6. An equipment fault diagnosis method based on federated self-supervised learning according to claim 5, characterized in that, in the S311, the first data enhancement sample It is obtained by first adding Gaussian noise to each piece of unlabeled data, and then scaling the noise-added data; the second data enhancement sample/> It is obtained by first smoothly warping the interval of time steps of each piece of unlabeled data and then applying noise. 7. An equipment fault diagnosis method based on federated self-supervised learning according to claim 4, characterized in that, during the training process, the local feature extractor outputs a first feature matrix and the first characteristic matrix/> Use the gradient descent algorithm to optimize the weight of the feature extractor. The loss function includes the first loss function and the second loss function. The calculation formula of the first loss function loss1 is as follows: Among them, N is the batch size, α and β are the first and second hyperparameters respectively, is the indicator function; i and j represent the first and second index of the sample in the batch respectively; l _c is the context contrast loss function, l _t is the time contrast loss function/> Represents the first clipping feature matrix, /> represents the second cropping feature matrix, s1 and s2 represent the first and second starting positions respectively, e1 and e2 represent the first and second ending positions respectively; T represents the length of the time dimension of the feature matrix;/> Represents the second clipping feature matrix/> Characteristics of sample i at time step t′,/> Represents the first clipping feature matrix/> Characteristics of sample i at time step t,/> Represents the first clipping feature matrix/> Sample i in ,/> Represents the second clipping feature matrix/> Sample i in ,/> Represents the second clipping feature matrix/> Sample j in The calculation formula of the second loss function loss2 is as follows: in, Respectively represent the first feature matrix extracted by the local feature extractor of the k-th client in the r-th round/> and the first characteristic matrix/> Respectively represent the first feature matrix extracted by the local feature extractor of the p-th round of the k-th client/> and the first characteristic matrix/> Represents the features extracted by each client using the global feature extractor weights received from the server at round r of training. 8. A device fault diagnosis method based on federated self-supervised learning according to claim 4, characterized in that, in the S32, the server uses a weighted average method to aggregate the local feature extractor weights of all clients, and the calculation formula is as follows: Among them, k is the client index, |D|, | ^Dk | represent the data volume of the global client and the data volume of the k-th client respectively, θ ^G and θ ^k represent the weight of the global feature extractor and the k-th client respectively. Local feature extractor weights at the end. 9. A device fault diagnosis method based on federated self-supervised learning according to claim 1, characterized in that, in the S4, each client performs the following steps: S51: Input the label data set into the local feature extractor with updated weights to obtain the feature matrix data set; S52: Use the feature matrix data set as the input of the support vector machine and train the support vector machine to obtain a client classifier. 10. A device fault diagnosis method based on federated self-supervised learning according to claim 1, characterized in that, in the S5, a sliding window is used to perform signal segmentation on the sensor data of the device to be diagnosed and a maximum-minimum method is used. After normalization processing, it is input into the fault diagnosis model of the corresponding client.