Bearing fault diagnosis method based on graph neural network self-supervised domain adaptation

Abstract

The invention discloses a bearing fault diagnosis method based on graph neural network self-supervised domain adaptation, aims at the field of domain self-adaptive fault diagnosis under variable working conditions, and aims at solving the problems of large data feature distribution difference under different working conditions, long migration learning domain adversarial training time and the like. The method comprises a graph sample construction module and a graph neural network maximum classifier difference field adaptation training module, a feature extractor and a double classifier are constructed by adopting a graph neural network, and in the training process, firstly, source domain classification and preliminary feature alignment of a source domain and a target domain are carried out; and then maximizing the output difference of the two classifiers to update the classifiers, freezing the parameters of the feature extractor to obtain the judgment boundary of the target domain, finally minimizing the output difference of the two classifiers to update the extractor to reduce the feature distance between the source domain and the target domain, and realizing cross-working-condition domain adaptation through adversarial training. A PU data set is adopted in an experiment, and compared with other domain adaptive fault diagnosis methods under variable working conditions, the method has remarkable advantages in diagnosis accuracy and model training time.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical equipment component fault diagnosis, relates to bearing fault diagnosis under variable working conditions, and particularly relates to a bearing fault diagnosis method based on graph neural network self-supervised domain adaptation. Background Art

[0002] With the continuous development of industrial technology and the improvement of automation levels, the demand for the production of various types of mechanical equipment and their parts is growing. Bearings, as key core components in mechanical systems, have a performance that is directly related to the stability and safety of the entire machine. However, actual operating conditions are complex and changeable, and bearing failures are frequent, posing a serious threat to the normal operation of the equipment. Traditional fault diagnosis methods rely heavily on manual experience and signal processing techniques, making it difficult to fully extract and characterize fault features. They are susceptible to noise interference when operating conditions vary significantly, and diagnostic accuracy and real-time performance cannot meet requirements. In recent years, the rapid development of artificial intelligence and deep learning technologies has provided new insights for fault diagnosis. Their automatic feature extraction and learning capabilities have, to a certain extent, compensated for the shortcomings of traditional methods. However, deep learning models still face challenges in migrating across operating conditions, such as the need for large amounts of data annotation, significant differences in feature distributions under different operating conditions, and lengthy adversarial domain adaptation training times. These issues severely restrict their widespread application in industrial settings.

[0003] Therefore, in order to address the deficiencies in the prior art, the present invention proposes a bearing fault diagnosis method based on self-supervised domain adaptation of graph neural networks. This method constructs graph samples, uses graph neural networks to build feature extractors and dual classifiers, first constructs a clear discrimination boundary by maximizing the judgment difference between the outputs of the two classifiers, and then reversely minimizes the classifier difference, and uses adversarial training to achieve effective matching of feature distributions across working domains, thereby overcoming the problem of inconsistent data distribution caused by changes in working conditions. At the same time, the use of a self-supervised strategy reduces dependence on a large amount of labeled data and shortens the training time of transfer learning, significantly improving the accuracy and real-time performance of fault diagnosis, and providing strong technical support for equipment safety management and intelligent manufacturing in industrial production. Summary of the Invention

[0004] This paper aims to provide a bearing fault diagnosis method and data processing flow based on self-supervised domain adaptation of a graph neural network. This method primarily addresses the significant differences in bearing data feature distribution under different operating conditions and the excessively long domain adversarial training time required for traditional transfer learning. It achieves efficient feature extraction, graph-structured self-supervised learning, and cross-domain adaptability, significantly improving fault diagnosis accuracy and training efficiency.

[0005] The method includes: graph sample construction of bearing dataset, graph neural network model building and domain adaptation training strategy;

[0006] Graph Sample Construction: The preprocessed data is segmented using random windows. The K-nearest neighbor algorithm is used to construct graph samples. The Euclidean distance is used to measure the edge weights between nodes. A graph is generated based on each window. The features of each graph node are represented by the normalized bearing signal features within the window, and the edge features are the weights calculated between nodes. The resulting graph data structure includes node features, edge indexes, and edge features.

[0007] Graph Neural Network Model Construction: This paper proposes an improved graph neural network architecture, comprising the following modules: Feature Extractor: This includes two simplified graph convolutional layers, each followed by batch normalization and an activation function. Dual Classifier Architecture: The two classifiers respectively contain a fully connected layer, a batch normalization layer, an activation function, a fully connected output layer, and a softmax output layer, enabling self-supervised differential classification learning.

[0008] Domain adaptation training strategy: Training includes three steps: Step 1: Use source domain and target domain data for forward propagation at the same time, calculate the source domain classification loss and domain alignment loss, and optimize the feature extractor and classifier parameters; Step 2: Maximize the difference between the two classifier outputs, freeze the feature extractor, and only update the classifier parameters; Step 3: Minimize the difference between the two classifier outputs, optimize the feature extractor parameters, and achieve cross-domain feature alignment.

[0009] Furthermore, the graph sample construction, graph neural network model building and domain adaptation training strategy of the bearing dataset specifically include:

[0010] To construct the graph sample for the bearing dataset, bearing vibration signal data was first collected and normalized to uniformly range within the [0, 1] range. The signal was then segmented and sampled, with each sample length set to 1024 and the number of samples set to 500. Random sampling was used. Furthermore, the time-domain signal was converted to the frequency domain. After dataset preprocessing, graph sample construction was performed. The K-nearest neighbor (KNN) algorithm was employed, using the Chebyshev distance to measure node distances. KNN node neighbors were determined, and for each node, the five nearest neighbor nodes were found to obtain the neighbor index. The Euclidean distance from a node to its neighbors was calculated, and the inter-node distances were converted using a Gaussian kernel function to obtain edge weights. Ten signal samples were grouped together to construct a graph. Each graph node was characterized by the normalized signal data, and graph edges were determined by the edge index and edge weights calculated using the aforementioned method. This resulted in the final graph structure.

[0011] The graph neural network model is constructed. The network model includes two modules: a feature extractor module and a dual classifier module. The feature extractor contains two layers of simplified graph convolutional layers (SGConv) and batch normalization (BatchNorm) layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and the convolution output feature dimension is 1024. The final output feature dimension of the feature extractor module is 1024. The node features after feature extraction are classified by two classifiers in turn. The structure of each classifier specifically includes a fully connected linear layer, a batch normalization layer, a ReLU activation function, a fully connected output layer, and a Softmax output layer. The output of each classifier is the category prediction probability distribution of the node. The above-mentioned detailed convolutional layers, batch normalization layers, linear classification layers, and activation functions constitute a complete graph neural network model, realizing efficient node feature extraction and classification prediction functions.

[0012] The domain adaptation training strategy involves the following steps: initial feature alignment and classifier training. In this step, forward propagation is performed using source and target domain data to optimize the parameters of the feature extractor and the two classifiers. The loss function consists of the source domain classification loss and the domain alignment loss (MMD loss). In the second step, the classifier difference maximization step, the feature extractor parameters are frozen, and only the two classifier parameters are optimized to maximize the difference between the two classifier outputs. The loss function consists of the following: source domain cross entropy classification loss, target domain dual classifier output root mean square error (RMSE) maximum difference loss, and target domain dual classifier output KL_CDD maximum difference loss. In the third step, the classifier difference minimization and domain feature alignment step, the parameters of the two classifiers are frozen, and only the feature extractor parameters are optimized with the goal of minimizing the difference between the target domain data and the dual classifier outputs. Four gradient optimizations are performed consecutively, with each optimization continuing the parameter update based on the parameters updated last. During each optimization process, the specific loss functions include the source domain cross-entropy classification loss, the target domain dual classifier output root mean square error (RMSE) difference minimization loss, the S_KLD difference minimization loss, and the target domain classification prediction entropy minimization loss. The learning rate dynamic adjustment strategy uses an adaptive learning rate scheduler with an initial learning rate of 0.001 and 100 epochs of training. The Adam optimizer is used with a batch size of 32 and a maximum number of iterations of 100. Through dynamic weight adjustment and an adaptive learning rate mechanism, the stability and convergence efficiency of domain adaptation training are improved.

[0013] Compared with the existing technology, the present invention provides a bearing fault diagnosis method and data processing flow based on self-supervised domain adaptation of graph neural networks. The advantages and technical effects of the present invention are as follows: the present invention realizes efficient domain adaptation training of cross-domain bearing data through an improved graph neural network structure and a maximum classifier difference strategy, effectively shortens the model training time and improves the accuracy of bearing fault diagnosis under variable working conditions, significantly improves the model generalization ability, and provides strong technical support for the intelligent operation and maintenance of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0015] Figure 1 A flowchart of the bearing fault diagnosis method based on graph neural network self-supervised domain adaptation provided by the present invention;

[0016] Figure 2 A network structure diagram of the bearing fault diagnosis method based on graph neural network self-supervised domain adaptation provided by the present invention;

[0017] Figure 3 The confusion matrix diagram of the bearing fault diagnosis method based on graph neural network self-supervised domain adaptation in the PU dataset source domain is a speed of 1500 rpm, a load torque of 0.7 Nm and a radial force of 1000 N, and the target domain is a speed of 1500 rpm, a load torque of 0.1 Nm and a radial force of 1000 N.

[0018] Figure 4 The bearing fault diagnosis method based on graph neural network self-supervised domain adaptation of the present invention is a t-SNE cluster diagram in which the source domain of the PU dataset is a speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N, and the target domain is a speed of 1500 rpm, a load torque of 0.1 Nm, and a radial force of 1000 N. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0020] Example 1

[0021] The present invention provides a bearing fault diagnosis method and system based on graph neural network self-supervised domain adaptation. Figure 1As shown in the figure, the main process of the bearing fault diagnosis method of the present invention includes: (1) raw data preprocessing; (2) graph sample construction; (3) network model; (4) loss function processing; (5) model training and testing; (6) model result comparison. The following introduces the bearing fault diagnosis method of the graph neural network self-supervised domain adaptation with the flowchart.

[0022] This example uses a bearing dataset from Paderborn University (PU) in Germany. This dataset includes four operating conditions, as shown in Table 1. Each operating condition includes one healthy state, seven outer ring fault states with varying degrees of artificial damage, and five inner ring fault states with varying degrees of artificial damage, for a total of 13 state categories. The sampling frequency is 64 kHz.

[0023] Table 1. Different working conditions of PU bearing dataset

[0024]

[0025] The above-mentioned step (1) raw data preprocessing specifically includes: firstly loading the raw vibration signal stored in the MAT file from a predefined file path, and then normalizing the vibration signal x to obtain x Normalization .

[0026]

[0027] Among them, x is the original data, x Normalization is the normalized data, min(x) and max(x) are the minimum and maximum values ​​of the original data respectively; next, the normalized signal is randomly intercepted into groups of 1024 sampling points as a single sample sequence, and converted into frequency domain data through fast Fourier transform FFT. The frequency domain conversion formula is:

[0028]

[0029] Where X(k) represents the amplitude of the signal at the kth frequency component, x(n) is the value of the original discrete signal at the nth sampling point, n ranges from 0 to L-1, where L is the length of the signal segment (number of sampling points), e is the base of the natural logarithm, and j is the imaginary unit. Represents the phase accumulation of the discrete Fourier transform (DFT) at the kth frequency component. Since only the positive frequency component needs to be retained for real-valued signals, the value range of k is After the transformation, to obtain the amplitude of each frequency component, it is necessary to take the absolute value |·| of the complex result and then divide it by the signal length L to complete the amplitude normalization. Finally, a sufficient number of 500 fault types were collected for each fault type to prepare the preprocessed data set for the next step of graph sample construction.

[0030] The above step (2) graph sample construction specifically includes: firstly, KNN classification is performed. In the graph sample construction, the distance between data points is first calculated using the K nearest neighbor algorithm. Suppose there is a data set χ={X1,X2,…,X n}, where each sample X i For each sample after preprocessing in the first step, the number of data points is 1024. For any sample to be classified X∈χ, calculate its difference from other samples X i The distance between them is measured using Chebyshev distance, and the formula is as follows:

[0031] d Chebyshev (X,X i )=max|XX i | (3)

[0032] Among them, X and X i Represents two samples respectively, the distance value d Chebyshev (X,X i ) represents the similarity between two samples. Next, the neighbor weight of each sample is calculated. In the KNN algorithm, after finding the k=5 nearest neighbors of the sample to be processed X, the influence of each neighbor on the current sample, i.e., the weight, needs to be calculated. In order to calculate the similarity strength between nodes, the Gaussian kernel function is used to convert the distance into a weight. For each neighbor X i , which is consistent with the sample X i The Euclidean distance between can be expressed as:

[0033]

[0034] Among them, x j and x i,j Represents sample X and neighbor sample X respectively i The value of the jth data point, L is the number of data points for each sample is 1024, the distance value d Euclidean (X,X i ) represents the similarity between the sample and its neighbor samples. Then, calculate the neighbor weight ω i , using the Gaussian kernel function, the form of the function is:

[0035]

[0036] where d i For samples X and X i The Euclidean distance between its i-th neighbors, σ is a parameter that controls the width of the Gaussian kernel. iRepresents the similarity between samples. The smaller the distance, the greater the similarity and the greater the weight. The weights of k=5 neighbors are calculated, and then the edge set of the graph is constructed. Let the nodes of the graph be X1, X2,…, X n , each edge represents the similarity strength between two nodes. For each node X i An edge is formed between it and its k=5 nearest neighbor nodes, and the weight of the edge is the weight ω calculated above i Specifically, the edge set ε of a graph can be represented as a two-dimensional array, where each row represents an edge:

[0037]

[0038] where e i =(X i ,X ik ,ω i ) represents the relationship between the i-th node and its k-th neighbor node X ik The edge between i is the weight between a node and its neighboring nodes. Finally, the graph sample is constructed by calculating the edge set ε and edge weight ω between the samples. i , converting each sample into a graph node and establishing connections based on the similarity between nodes. Each sample will be regarded as a node in the graph, and the calculated edges and edge weights will be used as input to the graph neural network. Specifically, the graph sample can be represented as a graph G:

[0039] G=(υ,ε,ω) (8)

[0040] where υ={X1,X2,…,X N} is the set of nodes in the graph, ε={(x i ,x j )} is the edge set in the graph, ω={(ω ij )} is the edge weight set, indicating that each edge (x i ,x j ) similarity.

[0041] Step (3) of the above mentioned network model construction: Figure 2As shown, the present invention uses a simplified graph convolutional network (SGCN), and the specific structure is as follows: first, a feature extractor module is defined, which includes two layers of simplified graph convolution (SGConv) and corresponding batch normalization (BatchNorm) and activation functions. The input feature dimension of the first layer of SGConv is the feature number feature of the input data, and the output feature dimension is 1024; then, a batch normalization layer (BatchNorm) with a channel number of 1024 and a ReLU activation function are used for nonlinear mapping processing. Then, the second layer of SGConv is entered, the input feature dimension of this layer is 1024, and the output dimension is also 1024, and feature extraction is completed again through the batch normalization layer and the ReLU activation function to obtain a feature representation with a final dimension of 1024. Two parallel classifier modules are connected after the feature extraction module, namely Classifier1 and Classifier2. The two classifiers have the same structure: the first layer is a linear fully connected layer with an input dimension of 1024 and an output dimension of 128; then it passes through a BatchNorm1d layer (input and output dimensions are both 128) and a ReLU activation function; then it passes through another linear fully connected layer to reduce the 128-dimensional features to the final required output channel dimension out_channel, that is, the number of bearing states; finally, the features are normalized using the Softmax function (along the first dimension) to obtain the category probability distribution of each node. During the forward propagation process of the above network model, the input is the node features (feature dimension), edge index and edge weight of the graph data, and graph convolution operations are performed through two layers of SGConv. Batch normalization and ReLU activation are performed after each convolution operation, and the output dimension is always maintained at 1024. The final 1024-dimensional features extracted are sent to the Classifier1 and Classifier2 classifiers for classification prediction. The output dimension of each classifier is out_channel, and two probability distribution vectors with the dimension of out_channel are generated as the final output of the network model.

[0042] In step (4) above, if Figure 2 As shown in Figure 1, the present invention designs and adopts a variety of loss functions based on the characteristics of source and target domain data to optimize the model training process. Source domain classification loss: For source domain data, the cross entropy loss function is used to measure the difference between the model prediction result and the true label. Let the true label of the source domain sample be y i , the predicted probability distribution of the model output is Then the cross entropy loss is defined as:

[0043]

[0044] where N Sis the number of source domain samples, C is the number of categories, y i,c represents the true label of the i-th sample in the c-th category, is the corresponding prediction probability. Domain adaptation loss: To reduce the difference between the feature distributions of the source domain and the target domain, a domain adaptation loss based on Maximum Mean Discrepancy (MMD) is used. Let the mean of the source domain feature be μ S , the target domain feature mean is μ T , then the MMD loss is defined as:

[0045]

[0046] where N S and N T are the number of samples in the source domain and target domain respectively, φ(·) is the feature mapping function of the network model constructed in step 3, and Represent the source domain and target domain samples respectively. Target domain consistency loss (Root Mean Square Error, RMSE): To ensure the consistency of the outputs of different classifiers in the target domain, the root mean square error (RMSE) loss is used, which is defined as:

[0047]

[0048] in, and They represent the prediction results of the i-th target domain sample by the two classifiers C1 and C2 respectively. The entropy loss calculates the entropy of the average predicted probability distribution of the target domain samples. The specific formula is:

[0049]

[0050] Where is a positive performance number that tends to zero. In actual calculations, it is used to prevent numerical calculation anomalies. The target domain distribution difference loss, the two-way symmetric KL divergence (S-KLD), can be expressed as:

[0051]

[0052] in and They represent the probability values ​​predicted by two different classifiers in the model for the same target domain sample (i-th sample) in category j. Dynamic loss weight adjustment: In order to balance the impact of each loss term on model training, a dynamic loss weight adjustment mechanism is introduced. Let the source domain classification loss be The target domain consistency loss is L RMSE , the target domain distribution difference loss is L S_KLD, the entropy loss is L Entropy , the corresponding weight coefficients α, η and β are dynamically adjusted according to the following formula:

[0053]

[0054] Among them, λ1, λ2 and λ3 are preset hyperparameters, which are 0.1, 0.01 and 0.01 respectively, used to control the relative importance of each loss term. The final comprehensive loss function is:

[0055]

[0056] Where i represents the second and third steps in the training process. The coefficient of the second step is -1, which means that the target domain difference loss is amplified, and the coefficient of the third step is 1, which means that the target domain difference loss is reduced.

[0057] In step (5) above, the model domain adaptation training strategy is as follows: Figure 2 As shown, according to the loss function processing in step (4), the source domain classification loss is set to Reduce the difference loss between the source domain and target domain feature distribution to L DAN , the target domain consistency loss is L RMSE , the target domain distribution difference loss is L S_KLD , the entropy loss is L Entropy The specific process of the model domain adaptation training strategy is as follows: the first step is preliminary domain feature alignment and classifier training, and the loss function consists of and L DAN ; The second step: maximize the difference between the classifier target domain output, freeze the feature extractor parameters, and only update the two classifiers. The loss function consists of L RMSE 、L S_KLD and L Entropy Step 3: Minimize the difference between the target domain outputs of the classifiers. Freeze the parameters of the two classifiers and optimize only the parameters of the feature extractor. Targeting to minimize the difference between the target domain data and the outputs of the two classifiers, perform four gradient optimizations in succession. Each optimization is based on the parameters updated last time. In each optimization process, the specific loss function consists of L RMSE 、L S_KLD and L Entropy .

[0058] Comparison of model results in step (6) above: Figure 3 and Figure 4 As shown in the figure, after the network model proposed in the present invention undergoes domain adaptation training, the trained model parameters are saved and applied to the target domain test set for verification, and the obtained results are shown in the figure. Figure 3 This is the confusion matrix diagram corresponding to the target domain test set, where the vertical axis and horizontal axis represent the actual label and predicted label of the sample respectively. It can be seen intuitively from the confusion matrix that the method of the present invention has high prediction accuracy on the target domain data and shows fewer misclassifications. Figure 4 The figure shows the t-SNE cluster visualization of the target domain feature representation. From the cluster diagram, we can observe that after training with the domain adaptation method of the present invention, samples of the same category in the target domain in the feature space are clearly clustered together, while the distance between samples of different categories is more obvious. The model significantly enhances the distinguishability of features and the ability to discriminate categories through the domain adaptation method, thereby effectively improving the accuracy of target domain classification. Figure 3 and Figure 4 As a result, the method proposed in the present invention effectively realizes the adaptation of the model to the target domain data and improves the recognition accuracy in practical applications.

[0059] To illustrate the training time advantage of our proposed bearing fault diagnosis method based on self-supervised domain adaptation using a graph neural network, we compared it with other classic convolutional neural network domain adaptation methods: CNN-MCD and CNN-BCD. The runtimes in the table refer to the time required for 100 training iterations of each model.

[0060] Table 2. Comparison of training time for different model domain adaptation methods

[0061]

[0062] It can be seen from Table 2 that the bearing fault diagnosis method based on graph neural network self-supervised domain adaptation proposed in the present invention has a shorter running time than the classic convolutional neural network based domain adaptation model.

Claims

1. A bearing fault diagnosis method based on graph neural network self-supervised domain adaptation, characterized in that: include: Graph sample construction is used to represent bearing signal characteristics in the graph domain; Graph neural network model construction: used to extract features and perform classification prediction on the graph data, specifically including a feature extractor and a dual classifier structure; The domain adaptation training strategy is used to train the graph neural network model, including the following steps: Step 1: Feature alignment and initial classification training. Use source domain data and target domain data for simultaneous forward propagation, calculate the classification loss of the source domain and the MMD loss of domain alignment, and optimize the parameters of the feature extractor and classifier. Step 2: Maximize the difference between classifier outputs, freeze the parameters of the feature extractor, and only optimize the parameters of the two classifiers to maximize the difference between the target domain data and the outputs of the two classifiers. The specific loss functions include the source domain cross entropy classification loss, the root mean square error (RMSE) difference maximization loss of the target domain two classifier outputs, the S_KLD difference maximization loss, and the entropy minimization loss of the target domain classification prediction; Step 3: Minimize the difference in classifier outputs, freeze the parameters of the two classifiers, and only optimize the parameters of the feature extractor. The goal is to minimize the difference between the target domain data and the outputs of the two classifiers. Perform gradient optimization four times in a row. Each optimization continues to update the parameters based on the last updated parameters. In each optimization process, the specific loss functions include the source domain cross entropy classification loss, the root mean square error (RMSE) difference minimization loss of the target domain two classifier outputs, the S_KLD difference minimization loss, and the entropy minimization loss of the target domain classification prediction.

2. The method according to claim 1, wherein: include: S1: Graph sample construction: After dataset preprocessing, graph sample construction is performed. The K-nearest neighbor (KNN) algorithm is used. The Chebyshev distance is used to measure the distance between nodes. The KNN node neighbors are determined. For each node, the five nearest neighbor nodes are found to obtain the neighbor index of each node. Calculate the Euclidean distance from a node to its neighboring nodes, convert the distance between nodes using the Gaussian kernel function to obtain edge weights, group every 10 signal samples, and construct a graph. Each graph node is characterized by the normalized signal data, and the graph edges are determined by the edge index and edge weight calculated by the above method, forming the final graph structure data; S2: Network model construction: A simplified graph convolutional network (SGCN) is used, and the specific structure is as follows: First, a feature extractor module is defined, which includes two layers of simplified graph convolution (SGConv) and corresponding batch normalization (BatchNorm) and activation function; the input feature dimension of the first layer of SGConv is the feature number feature of the input data, and the output feature dimension is 1024; then a batch normalization layer (BatchNorm) with a channel number of 1024 and a ReLU activation function are used for nonlinear mapping processing; then the second layer of SGConv is entered, the input feature dimension of this layer is 1024, and the output dimension is also 1024, and the feature extraction is completed again after the batch normalization layer and the ReLU activation function, and the final feature representation with a dimension of 1024 is obtained; after the feature extraction module, two parallel classifier modules are connected, namely Classifier1 and Classifier2; the two classifiers have the same structure, and the specific structure is: the first layer is a linear fully connected layer with an input dimension of 1024 and an output dimension of 12 8; then pass through the BatchNorm1d layer (input and output dimensions are both 128) and the ReLU activation function; then pass through a linear fully connected layer to reduce the 128-dimensional feature to the final required output channel dimension out_channel, that is, the number of bearing states; finally, use the Softmax function (along the first dimension) to normalize the features to obtain the category probability distribution of each node; in the forward propagation process of the above network model, the input is the node features (dimension is feature), edge index and edge weight of the graph data, and the graph convolution operation is performed through two layers of SGConv. After each layer of convolution operation, batch normalization and ReLU activation processing are performed, and the output dimension is always maintained at 1024; the 1024-dimensional features finally extracted are sent to the two classifiers Classifier1 and Classifier2 for classification prediction; the output dimension of each classifier is out_channel, and two probability distribution vectors with the dimension of out_channel are generated respectively as the final output of the network model; S3: Domain adaptation training strategy: The training loss function consists of the source domain classification loss. For the source domain data, the cross entropy loss function is used to measure the difference between the model prediction result and the true label; let the true label of the source domain sample be y i , the predicted probability distribution of the model output is Then the cross entropy loss is defined as: where N S is the number of source domain samples, C is the number of categories, y i,c represents the true label of the i-th sample in the c-th category, is the corresponding prediction probability; domain adaptation loss: to reduce the difference between the feature distribution of the source domain and the target domain, the domain adaptation loss based on the maximum mean discrepancy (MMD) is adopted; the mean of the source domain feature is μ S , the target domain feature mean is μ T , then the MMD loss is defined as: where N S and N T are the number of samples in the source domain and the target domain respectively, φ(·) is the feature mapping function of the network model constructed by the method of the present invention, and Represent the source domain and target domain samples respectively; Target domain consistency loss (Root Mean Square Error, RMSE): To ensure the consistency of the outputs of different classifiers in the target domain, the root mean square error (RMSE) loss is used, which is defined as: in, and They represent the prediction results of the i-th target domain sample by the two classifiers C1 and C2 respectively; the entropy loss calculates the entropy of the average predicted probability distribution of the target domain samples. The specific formula is: Where ε is a very small positive number that tends to zero. In actual calculations, 1e-6 is used to prevent numerical calculation anomalies. The target domain distribution difference loss, the two-way symmetric KL divergence (S-KLD), can be expressed as: in and They represent the probability values ​​predicted by two different classifiers in the model for the same target domain sample (i-th sample) in category j; Dynamic loss weight adjustment: In order to balance the impact of each loss term on model training, a dynamic loss weight adjustment mechanism is introduced; let the source domain classification loss be The target domain consistency loss is L RMSE , the target domain distribution difference loss is L S_KLD , the entropy loss is L Entropy , the corresponding weight coefficients α, η and β are dynamically adjusted according to the following formula: Among them, λ1=0.1, λ2=0.01 and λ3=0.01 are preset hyperparameters used to control the relative importance of each loss term; the final comprehensive loss function is: Where i is the second and third steps in the training step; the coefficient of the second step is -1, which means amplifying the target domain difference loss, and the coefficient of the third step is 1, which means reducing the target domain difference loss; according to the loss function composition, let the source domain classification loss be Reduce the difference loss between the source domain and target domain feature distribution to L DAN , the target domain consistency loss is L RMSE , the target domain distribution difference loss is L S_KLD , the entropy loss is L Entropy The specific process of the model domain adaptation training strategy is as follows: the first step is preliminary domain feature alignment and classifier training, and the loss function consists of and L DAN ; The second step: maximize the difference between the classifier target domain output, freeze the feature extractor parameters, and only update the parameters of the two classifiers. The loss function consists of L RMSE , L S_KLD and L Entropy ; The third step: minimize the difference in the target domain output of the classifier, freeze the parameters of the two classifiers, and only optimize the parameters of the feature extractor. The goal is to minimize the difference between the target domain data and the output of the two classifiers. Perform gradient optimization four times in a row. Each optimization is based on the parameters of the last update. In each optimization process, the specific loss function consists of L RMSE , L S_KLD and L Entropy .

Images