Device fault diagnosis method based on adversarial transfer learning and class balance loss

Abstract

The application discloses a kind of based on the equipment fault diagnosis method of class balance loss of adversarial transfer learning, steps include:1 to the original signal data collected is preprocessed and constructs source domain and target domain sample set;2 constructs class non-equilibrium adversarial transfer learning network;3 class non-equilibrium adversarial transfer learning network model is trained using the way of joint optimization, and model parameters are optimized;4 test sample in target domain sample set is input into the model trained, and the fault diagnosis result of equipment is obtained.The application can realize the fault diagnosis of oil and gas pipeline under different working conditions, and can well deal with the class non-equilibrium problem caused by less fault state under mechanical actual working condition, so as to improve the fault diagnosis effect and generalization performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fault diagnosis, and particularly relates to a device fault diagnosis method based on an adversarial transfer learning and a class balance loss. BACKGROUND

[0002] Mechanical equipment is an important part of modern industrial production and is widely used in various fields of modern industry, such as oil and natural gas exploitation, general equipment manufacturing, etc. Due to the complication and precision of the internal structure of mechanical equipment, the difficulty of equipment state monitoring and equipment management is greatly improved. Once mechanical equipment fails, it may cause great resource loss and economic loss, or even cause catastrophic accidents and threaten life safety. Therefore, timely detection of equipment failure and positioning and diagnosis of the failure have important practical and economic significance for ensuring the safe operation of the equipment.

[0003] Currently, there are two methods for diagnosing equipment failure: a method based on a physical model and a data-driven method. The method based on a physical model analyzes the failure mechanism of the equipment, combines prior knowledge and the running state of the equipment to establish a mathematical model, and then diagnoses the health state of the equipment. However, due to the complexity of the working environment and the variability of the working state of the equipment, it is difficult to establish an accurate physical model of the equipment. Compared with the method based on a physical model, the data-driven diagnosis method relies on sensor signal data and uses statistical and machine learning methods to establish a mapping relationship between the state monitoring signal and the failure state, and has certain advantages in terms of precision and generalization ability. As one of the important technologies of data-driven, deep learning has strong ability to automatically extract features from sensor data and is widely used in the field of fault diagnosis.

[0004] However, on the one hand, the good performance of most deep learning methods is usually limited by two assumptions: the test data and the training data need to satisfy the independent and identically distributed assumption, and there are sufficient labeled fault samples for the task to be diagnosed. However, in actual working scenarios, the distribution of equipment state monitoring data under different working conditions has great difference, and it is difficult to collect sufficient sample data with fault labels on the equipment to be detected. On the other hand, most current deep learning methods are based on the assumption that different health state data are balanced. However, in actual working conditions, most of the equipment is in a healthy running state, and the probability of failure is small, so most of the collected equipment running data is healthy state data, and the fault data is less. At the same time, affected by multiple factors of the working environment, the probability of occurrence of different types of equipment failure is different, resulting in a great difference in the amount of sample data of different fault types in the equipment state monitoring data. Therefore, the equipment state monitoring data has a serious imbalance phenomenon, which greatly affects the practical application of deep learning methods in pipeline fault diagnosis. SUMMARY

[0005] The application aims to provide a device fault diagnosis method based on adversarial transfer learning and class balance loss to eliminate the influence of the difference between data feature distributions under different working conditions and the data class imbalance problem on the device fault diagnosis model, thereby improving the accuracy of device fault diagnosis.

[0006] To achieve the above application purposes, the application adopts the following technical solutions:

[0007] The device fault diagnosis method based on domain adversarial transfer learning and class balance loss comprises the following steps:

[0008] Step 1, preprocessing the collected original signal data and constructing source domain and target domain sample sets:

[0009] Step 1.1, obtaining multi-sensor monitoring signal data of the device under a certain working condition and setting it as source domain data, obtaining multi-sensor monitoring signal data of the device under other working conditions and setting it as target domain data; and performing normalization processing on all monitoring signal data to obtain normalized source domain data and normalized target domain data;

[0010] Step 1.2, performing sample division on the normalized source domain data by using a sliding time window method to obtain a source domain sample set wherein N represents the number of source domain samples, is the i-th source domain sample, represents the source domain multi-sensor monitoring signal data in the i-th sliding window, represents the k-th source domain sensor monitoring signal data in the i-th sliding window, K represents the number of sensors, and l represents the size of the time window; y i is the corresponding health state label, y i ∈{1,2,…,C}; C represents the number of health state categories. is the corresponding domain label;

[0011] Step 1.3, performing sample division on the normalized target domain data according to the process of step 1.2 to obtain a target domain sample set wherein M represents the number of target domain samples, is the j-th target domain sample, represents the k-th target domain sensor monitoring signal data in the j-th sliding window; represents the corresponding domain label;

[0012] Step 2, constructing a class imbalance adversarial transfer learning network, comprising: a public feature extractor Pe , an imbalanced classifier I c , a domain discriminator D dc ;

[0013] Step 2.1, constructing a public feature extractor P e is composed of K branches containing W layers of depth separable convolution layers and a multi-sensor attention mechanism branch in sequence; each depth separable convolution layer is composed of a pointwise convolution layer and a depth convolution layer in sequence; the multi-sensor attention mechanism branch is composed of a global average pooling layer and two fully connected layers in sequence; and a ReLU activation function is connected after the first fully connected layer, and a Sigmoid activation function is connected after the second fully connected layer;

[0014] the i-th source domain sample is input into the public feature extractor P e , the source domain multi-sensor monitoring signal data in the sample is processed by the K depth separable convolution layer branches in sequence, and K sensor depth feature vectors are output in sequence; then the K sensor depth feature vectors are input into the sensor attention mechanism branch for processing, and the i-th source domain depth feature vector is obtained, wherein represents the depth feature vector of the k-th sensor monitoring signal in the i-th source domain sample;

[0015] the j-th target domain sample is input into the public feature extractor P e for processing, and the j-th target domain depth feature vector

[0016] Step 2.2, constructing a domain discriminator D dc is composed of a gradient reversal layer, an F-layer fully connected network and a 1-layer Softmax activation function in sequence; and a ReLU activation function is used as the activation function after each fully connected layer;

[0017] the i-th source domain depth feature vector is input into the domain discriminator D dc for domain recognition, and the domain discrimination result of is obtained

[0018] the j-th target domain depth feature vector is input into the domain discriminator D dc for domain recognition, and the domain discrimination result of is obtained

[0019] Step 2.3, constructing an imbalanced classifier I cThe F layer full connection network and the 1 layer Softmax activation function are sequentially composed; and the ReLU activation function is used as an activation function after each full connection layer;

[0020] The i-th source domain deep feature vector Input to the non-equilibrium classifier I c The health state is recognized, and The prediction result corresponding to the real sample label

[0021] Step 3, constructing a global objective function L composed of a class non-equilibrium focus loss function L cb And an adversarial domain self-adaptive learning loss function L dac The gradient descent method is used to train the class non-equilibrium adversarial transfer learning network, and the global objective function L is minimized until convergence, so as to obtain an optimal equipment fault diagnosis model, and then the target domain samples not participating in training are detected, so as to obtain the health state of the equipment.

[0022] The equipment fault diagnosis method based on the adversarial transfer learning and the class balance loss has the characteristics that the global objective function L is obtained as follows:

[0023] Step 3.1, a class non-equilibrium focus loss function L is established by using formula (1) cb :

[0024]

[0025] In formula (1), n s represents the number of a batch of samples from the normalized source domain data; I yi represents the number of samples with the same label as the i-th sample in the source domain sample; γ, β, δ represent three hyperparameters;

[0026] Step 3.2, an adversarial domain self-adaptive learning loss function L is established by using formula (2) dac :

[0027]

[0028] In formula (2), L d represents a supervised loss function; n t represents the number of a batch of samples from the normalized target domain data;

[0029] Step 3.3, a global objective function L is constructed by using formula (3):

[0030]

[0031] In formula (3), λd This represents the penalty coefficient.

[0032] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the device fault diagnosis method, and the processor is configured to execute the program stored in the memory.

[0033] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the device fault diagnosis method.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention utilizes deep learning methods to achieve end-to-end prediction of equipment health status from raw sensor data without requiring prior knowledge of pipeline operation. It solves the problem of complex data processing and feature extraction required by existing methods and has the advantages of simple operation and strong generalization.

[0036] 2. The method of the present invention utilizes an adversarial transfer learning strategy to fully extract the common features of equipment monitoring signal data under different working conditions, and adaptively eliminates the marginal distribution differences of monitoring signal data under different working conditions, thereby improving the generalization performance of equipment fault diagnosis.

[0037] 3. The method of this invention utilizes class-equalized focus loss, which fully considers the severe imbalance problem of equipment monitoring signal data, distinguishes the differences in the model training loss of samples of different health states of the equipment, and makes the model pay more attention to the small sample size of fault category data that is difficult to identify, thereby improving the accuracy of equipment fault diagnosis. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 This is a structural diagram of the non-equilibrium adversarial transfer learning network of the method of the present invention;

[0040] Figure 3 This is a structural diagram of the common feature extraction module of the present invention. Specific implementation methods

[0041] In this embodiment, as Figure 1 As shown, the process of a device fault diagnosis method based on adversarial transfer learning and class balance loss includes the following steps:

[0042] Step 1: Preprocess the acquired raw signal data and construct source and target domain sample sets:

[0043] Step 1.1, obtaining the monitoring signal data of the equipment under a certain working condition and setting it as the source domain data, obtaining the monitoring signal data of the equipment under other working conditions and setting it as the target domain data; and normalizing all the monitoring signal data to obtain the normalized source domain data and the normalized target domain data; in this embodiment, the monitoring data of the bearing under the working conditions of N15_M07_F10 and N09_M07_F10 (N represents the rotating speed, M represents the loaded torque, and F represents the radial force) in the Paderborn data set are selected as the source domain data, and the monitoring data of the bearing under the working conditions of N15_M01_F10 and N15_M07_F04 (N represents the rotating speed, M represents the loaded torque, and F represents the radial force) are selected as the target domain data;

[0044] Step 1.2, dividing the normalized source domain data into samples by using the sliding time window method to obtain the source domain sample set wherein, N represents the number of source domain samples, is the i-th source domain sample, represents the source domain multi-sensor monitoring signal data in the i-th sliding window, represents the k-th source domain sensor monitoring signal data in the i-th sliding window, K represents the number of sensors, and l represents the size of the time window. i is the corresponding health state label, y i ∈{1,2,…,C};C represents the number of health state categories. is the corresponding domain label; in this example, the size of the sliding window is uniformly set to 2048; is uniformly set to 0;

[0045] Step 1.3, dividing the normalized target domain data into samples according to the process of step 1.2 to obtain the target domain sample wherein, M represents the number of target domain samples, is the j-th target domain sample, represents the k-th target domain sensor monitoring signal data in the j-th sliding window. represents the corresponding domain label; in this example, the size of the sliding window is uniformly set to 2048; is uniformly set to 1;

[0046] Step 2, constructing a non-balanced adversarial transfer learning network, as shown in Figure 2 , which includes: a public feature extractor P e , a non-balanced classifier I c , and a domain discriminator D dc ;

[0047] Step 2.1, as shown in Figure 3 , a public feature extractor P e is constructed, which is composed of K branches containing W layers of depth separable convolution layers (in this example, K is set to 6 and W is set to 10) and a multi-sensor attention mechanism branch in sequence; each depth separable convolution layer is composed of a pointwise convolution layer (in this example, the convolution kernel size is 1x1) and a depth convolution layer (in this example, the convolution kernel size is 3x3 and the padding is set to 1) in sequence; the multi-sensor attention mechanism branch is composed of a global average pooling layer and two fully connected layers in sequence; and a ReLU activation function is connected after the first fully connected layer, and a Sigmoid activation function is connected after the second fully connected layer;

[0048] In order to obtain a rich feature representation of multi-sensor data, the i-th source domain sample is input into the public feature extractor P e , and the multi-sensor monitoring signal data is processed by K depth separable convolution layer branches in sequence, and K depth feature vectors of the monitoring signal data are output in sequence wherein represents the depth feature vector of the k-th sensor monitoring signal in the i-th source domain sample; and then the two are input into the channel attention mechanism branch for processing, and the global feature points of the multi-sensor depth features are obtained by using the global average pooling layer After splicing, the multi-sensor feature weight is calculated by using the fully connected layer and the feature weighted fusion is performed to reduce the redundancy of feature information, realize the fusion of multi-sensor information, and obtain the i-th source domain depth feature vector In this example, the dimension of which is 512;

[0049] The j-th target domain sample is input into the public feature extractor P e for processing, and the j-th target domain depth feature vector is obtained after multi-sensor feature extraction and fusion. In this example, the dimension of which is 512;

[0050] Step 2.2, a domain discriminator D dc is constructed, which is composed of a gradient reversal layer, an F-layer fully connected network (in this example, F is set to 3) with an input length of 512, and a layer of Softmax activation function in sequence; and a ReLU activation function is used as the activation function after each fully connected layer;

[0051] the i-th source domain depth feature vector input into the domain discriminator D dc after domain recognition, to obtain a domain discrimination result of

[0052] the jth target domain deep feature vector input into the domain discriminator D dc after domain recognition, to obtain a domain discrimination result of

[0053] Step 2.3, constructing an imbalanced classifier I c consisting of an F-layer fully connected network with an input length of 512 and a 1-layer Softmax activation function in sequence; and a ReLU activation function is used as the activation function after each fully connected layer;

[0054] the ith source domain deep feature vector input into the imbalanced classifier I c for health state recognition, to obtain a prediction result corresponding to the real sample label

[0055] Step 3, training of the class-imbalanced adversarial transfer learning network;

[0056] Step 3.1, introducing an adversarial domain self-adaptive learning strategy to enable the public feature extractor P e to learn the common domain distribution existing in the source domain and the target domain. Specifically, when the domain discriminator D dc tries to distinguish the source domain features and the target domain features as much as possible, the public feature extractor P e tries to make the generated deep features and unable to be recognized by the domain discriminator D dc . Therefore, the public feature extractor P e and the domain discriminator D dc form an adversarial relationship. A class-imbalanced focal loss function L cb is established using formula (1):

[0057]

[0058] In formula (1), n s represents the number of a batch of samples from the normalized source domain data (in this example, n s is set to 48); I yi ​​​represents the number of samples in the source domain samples with the same label as the i-th sample; γ, β, δ represent three hyperparameters; in this example, γ is set to 2, β is set to 0.2, and δ is set to 1;

[0059] Step 3.2, introduce a class imbalance learning strategy, so that the public feature extractor P e pays more attention to the feature learning of the minority class samples and the difficult-to-classify samples, thereby improving the recognition effect of the minority class samples on the imbalance classifier I c . The formula (2) is used to establish the adversarial domain self-adaptive learning loss function L dac :

[0060]

[0061] In formula (2), L d represents a supervised loss function (such as square loss, loss, cross-entropy loss, etc. In this example, cross-entropy loss is used); n t represents the number of samples from the normalized target domain data (in this example, n t is set to 16);

[0062] Step 3.3, sample imbalance learning and adversarial domain self-adaptive learning are performed by joint optimization, and formula (3) is used to construct a global objective function L:

[0063]

[0064] In formula (3), λ d represents a penalty coefficient, and in this example, λ d is set to 1.1;

[0065] Step 3.4, based on the global objective function L, the gradient descent method is used to train the class imbalance adversarial transfer learning network, and the global objective function L is minimized until convergence, thereby obtaining the optimal equipment fault diagnosis model. In this example, the training optimizer is selected as the adam optimizer, the learning rate is set to 0.01, and the total training number is set to 100 rounds; then, the target domain samples not participating in the training are detected, thereby obtaining the health status of the equipment.

[0066] In this embodiment, an electronic device includes a memory for storing a program supporting a processor to execute the above method, and a processor configured to execute the program stored in the memory.

[0067] In this embodiment, a computer readable storage medium has a computer program stored thereon, and the computer program is executed by a processor to perform the steps of the above method.

[0068] The following is a specific example of the experimental demonstration of the method of the present application, and the specific content is as follows:

[0069] 1. Data set:

[0070] The data used in the present application is a gearbox data set provided by the University of Paderborn in Germany, which includes 6 sensor detection variables. The training data set and the test data set in the experiment each include the running data of 8 bearings under different test conditions, wherein the training data set is the full life cycle data of the bearing, and the test data set is the data of the bearing in a period of time.

[0071] 2. Evaluation index:

[0072] The present application uses F1 score (F1-score) and geometric mean (Geometric mean, G-mean) as evaluation indexes. The calculation scores of F1-score and G-mean are as follows:

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] Among them, TP represents the number of true cases, TN represents the number of true negative cases, FP represents the number of false positive cases, FN represents the number of false negative cases, Precision represents the accuracy of model classification, Recall represents the recall rate of model classification, Sensitivity represents the sensitivity of model classification, and Specificity represents the specificity of model classification. The greater the value of F1-score and G-mean, the smaller the error of fault diagnosis, and the better the diagnostic performance of the model.

[0080] 3. Experimental process:

[0081] To verify the effectiveness of the proposed method, the present application is compared with other methods. The comparison methods include Multilayer Perceptron (MLP), Convolution Neural Network (CNN), Gate Recurrent Unit (GRU), Domain Adaptation Network (DAN), Domain Adversarial Neural Network (DANN), and the Class Imbalanced Domain Adversarial Transfer Learning Network (CIB-DATLN) proposed in this paper. In the experiment, the main structure of the MLP model is three fully connected layers and two ReLU activation functions. The main structure of the CNN model is a convolution layer, a pooling layer, and a fully connected layer. The GRU model selects the features of the last time step as input to the fully connected layer to predict the sample fault type. The DAN model uses a fully connected layer for feature extraction and uses the maximum mean difference learning domain invariance feature. The DANN model uses CNN and ReLU activation function for feature extraction, and uses a fully connected layer to predict the sample fault type, and realizes adversarial transfer learning through the gradient reversal layer (GRL). During the experiment, the present application selects 80% of the source domain labeled data samples and 40% of the target domain unlabeled data samples to train the model, uses 20% of the source domain labeled data samples and 10% of the target domain data samples as the verification data set to test the prediction effect of the model, and finally uses the remaining 50% of the target domain data as the test set, and evaluates the model effect on the test set.

[0082] 4. Experimental results

[0083] To verify the effectiveness of the proposed method, the present application is compared with other methods. The comparison methods include Multilayer Perceptron (MLP), Convolution Neural Network (CNN), Gate Recurrent Unit (GRU), Domain Adaptation Network (DAN), Domain Adversarial Neural Network (DANN), and the Class Imbalanced Domain Adversarial Transfer Learning Network (CIB-DATLN) proposed in this paper. In the experiment, the main structure of the MLP model is three fully connected layers and two ReLU activation functions. The main structure of the CNN model is a convolution layer, a pooling layer, and a fully connected layer. The GRU model selects the features of the last time step as input to the fully connected layer to predict the sample fault type. The DAN model uses a fully connected layer for feature extraction and uses the maximum mean difference learning domain invariance feature. The DANN model uses CNN and ReLU activation function for feature extraction, and uses a fully connected layer to predict the sample fault type, and realizes adversarial transfer learning through the gradient reversal layer (GRL). During the experiment, the present application selects 80% of the source domain labeled data samples and 40% of the target domain unlabeled data samples to train the model, uses 20% of the source domain labeled data samples and 10% of the target domain data samples as the verification data set to test the prediction effect of the model, and finally uses the remaining 50% of the target domain data as the test set, and evaluates the model effect on the test set.

[0084] Table 1 F1-score and G-mean experimental results

[0085] Method F1-score G-mean MLP 82.16 80.43 CNN 89.64 82.63 GRU 90.12 83.45 DAN 95.46 89.72 DANN 96.08 90.12 CIB-DATLN 98.45 92.41

[0086] As can be seen from Table 1, the method disclosed by the application is superior to the other five fault diagnosis methods in terms of F1-score and G-mean two evaluation indexes. As can be seen from Table 1, compared with other methods, the fault diagnosis accuracy of DAN and DANN with the transfer learning strategy is improved, and the fault diagnosis accuracy of the method of the application is further improved compared with DAN and DANN. The experimental results show that for bearing fault diagnosis, the method of the application can fully extract the field commonality features of gearbox data under different working conditions by using the adversarial transfer learning strategy, and can pay more attention to small sample category data difficult to identify by using the unbalanced focal loss, thereby improving the fault diagnosis effect of the gearbox.

Claims

1. A device fault diagnosis method based on domain adversarial transfer learning and class-balanced loss, characterized in that, Comprising the following steps: Step 1, preprocessing the collected original signal data and constructing source domain and target domain sample sets: Step 1.1, obtaining multi-sensor monitoring signal data of the equipment under certain working conditions and setting as source domain data, obtaining multi-sensor monitoring signal data of the equipment under other working conditions and setting as target domain data; and normalizing all monitoring signal data to obtain normalized source domain data and normalized target domain data; Step 1.

2. Sample division of the normalized source domain data using the sliding time window method to obtain the source domain sample set where N represents the number of source domain samples, is the i-th source domain sample, represents the source domain multi-sensor monitoring signal data in the i-th sliding window, represents the k-th source domain sensor monitoring signal data in the i-th sliding window, K represents the number of sensors, and l represents the size of the time window; y i is the corresponding health status label, y i ∈{1,2,…,C}; C represents the number of health status categories; is the corresponding domain label; Step 1.

3. Sample division is performed on the normalized target domain data according to the process of step 1.2 to obtain target domain samples wherein M represents the number of target domain samples, is the jth target domain sample, represents the kth target domain sensor monitoring signal data in the jth sliding window; represents the corresponding domain label; Step 2, constructing a non-equilibrium confrontation transfer learning network, comprising: a public feature extractor P e , a non-equilibrium classifier I c , a domain discriminator D dc ; Step 2.1, constructing the public feature extractor P e is composed of K branches containing W layers of depth separable convolution layers and a multi-sensor attention mechanism branch in sequence; each depth separable convolution layer is composed of a pointwise convolution layer and a depth convolution layer in sequence; the multi-sensor attention mechanism branch is composed of a global average pooling layer and two fully connected layers in sequence; and a ReLU activation function is connected after the first fully connected layer, and a Sigmoid activation function is connected after the second fully connected layer; The i-th source domain sample Input the common feature extractor P e In the method, the source domain multi-sensor monitoring signal data in the sample is sequentially subjected to W-layer convolution processing through K deep separable convolution layer branches, and K sensor deep feature vectors are correspondingly output The i-th source domain deep feature vector is obtained by inputting the sensor attention mechanism branch again for processing Wherein, The i-th source domain sample is a deep feature vector of the k-th sensor monitoring signal jth target domain sample input to the common feature extractor P e is processed to obtain a jth target domain deep feature vector Step 2.2, constructing the field discriminator D dc It is composed of a gradient reversal layer, an F layer full connection network and a Softmax activation function in sequence; and a ReLU activation function is used as the activation function after each full connection layer. the i-th source domain deep feature vector input to the domain discriminator D dc after domain identification, obtaining the domain discrimination result jth target domain deep feature vector input to the domain discriminator D dc after domain identification, obtaining the domain discrimination result Step 2.3, building the unbalanced classifier I c consists of F layers of fully connected networks and 1 layer of Softmax activation function in sequence; and a ReLU activation function is used as the activation function after each fully connected layer; i-th source domain deep feature vector input to the non-equilibrium classifier I c health state recognition, obtaining prediction result corresponding to the real sample label Step 3, constructing a non-equilibrium focus loss function L cb and an adversarial domain adaptive learning loss function L dac a global objective function L, and using gradient descent method to train the class non-equilibrium adversarial transfer learning network, and minimizing the global objective function L until convergence, thereby obtaining an optimal equipment fault diagnosis model, and then detecting the target domain samples not participating in the training, thereby obtaining the health state of the equipment.

2. The device fault diagnosis method based on domain adversarial transfer learning and class-balanced loss according to claim 1, characterized in that, The global objective function L is obtained as follows: Step 3.1, Establishing a class imbalanced focal loss function L with formula (1) cb : In formula (1), n s represents the number of a batch of samples from the normalized source domain data; I yi represents the number of samples in the source domain samples having the same label as the i-th sample; γ, β, δ represent three hyperparameters; Step 3.2, establish the adversarial domain adaptation learning loss function L with formula (2) dac : In formula (2), L d represents a supervised loss function; n t represents the number of a batch of samples from the normalized target domain data; Step 3.3, constructing the global objective function L using formula (3): In formula (3), λ d represents a penalty coefficient.

3. An electronic device comprising a memory and a processor, characterized in that The memory is used to store a program supporting the processor to execute the equipment fault diagnosis method of claim 1 or 2, and the processor is configured to execute the program stored in the memory.

4. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to execute the steps of the equipment fault diagnosis method of claim 1 or 2.

Images