A motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning

Abstract

The present invention discloses a motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning, comprising the following steps: first, obtaining multi-source signals of the motor in different health states and normalizing them, and establishing a sample set of original signal graphs; second, using graph samples in different health states of the motor to train the model, capturing the relationship features between cross-graph samples through a multi-head self-attention mechanism and a feedforward network, and combining the knowledge distillation technology to obtain an optimized cascade structure of a multi-layer graph isomorphism network; finally, using the trained model to test the samples to be diagnosed, and finally outputting the diagnosis results. The present invention enhances noise resistance by fusing multi-sensor signals, extracting cross-graph relationship features, and a new dynamic training feedback strategy, thereby significantly improving the accuracy and robustness of motor fault diagnosis. Experimental results show that the method proposed in the present invention is superior to the current mainstream fault diagnosis methods in both fault diagnosis accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of motor fault diagnosis, and specifically provides a novel motor fault diagnosis method based on graph isomorphism networks and cross-graph relationship learning. Background Art

[0002] Electric motors, due to their high efficiency, high power density, and excellent control performance, have been widely used in industrial drives, electric vehicles, and wind power generation. However, various faults inevitably occur during motor operation, such as bearing failures, stator winding faults, and rotor imbalance. If these faults are not diagnosed and addressed promptly, they can lead to degraded motor performance and even serious safety accidents. Therefore, developing efficient and accurate motor fault diagnosis methods is crucial to ensuring safe and reliable motor operation. In recent years, deep learning-based motor fault diagnosis methods have developed rapidly, particularly with advances in algorithms such as graph neural networks, which have significantly enhanced the models' data analysis capabilities.

[0003] Graph isomorphism networks (GINs), as an important branch of graph neural networks, have demonstrated significant advantages in the field of fault diagnosis due to their powerful feature extraction capabilities. GIN aggregates node features by combining multi-layer perceptrons and summation functions, and generates a global representation of the graph through a graph readout function. This design enables it to excel in tasks such as fault diagnosis that require precise identification of graph structures. However, most existing research methods are based on the analysis of single sensor signals, making it difficult to fully utilize the complementary information of multi-sensor data, and signal acquisition is easily affected by noise in actual industrial environments. At the same time, existing fault diagnosis methods based on graph isomorphism networks usually focus on a single sample, ignore the associations between graph samples, and cannot effectively capture the similarities or differences between samples of different fault types. This means that the diagnostic accuracy in motor fault diagnosis will decrease. Summary of the Invention

[0004] The purpose of the present invention is to address the above problems and provide a motor fault diagnosis method based on graph isomorphism networks and cross-graph relationship learning. The method includes acquiring and normalizing multi-sensor signals from the motor; obtaining the original graph representation of multi-source signal graph samples through a cascade structure of a multi-layer graph isomorphism network, making full use of the complementary information of multi-source data; combining a multi-head self-attention mechanism and a feedforward network to learn the cross-graph sample relationship features, learn the similarities and differences between samples of different fault types, and significantly improve the model's noise resistance; introducing technologies such as knowledge distillation to design a new dynamic feedback training strategy, transferring the relationship features of cross-graph samples to the cascade structure of the multi-layer graph isomorphism network to establish a dynamic feedback loop, so that the model can better understand the relationship features between cross-graph samples; taking graph samples of various fault states and corresponding motor fault labels as input, the model output is obtained and used as the motor fault diagnosis result. The present invention simultaneously considers the characteristics of multiple sensors in the same fault state and the relationship features between graph samples of different fault categories, and directly outputs the motor fault diagnosis result through an end-to-end training method, overcoming the limitations of existing methods in complex industrial environments, significantly improving diagnostic accuracy and robustness, and providing reliable support for motor fault diagnosis.

[0005] To achieve the above objectives, the present invention adopts a technical solution: a motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning, comprising the following steps:

[0006] Step 1: Obtain multi-source signals under different health conditions and establish an original signal sample set;

[0007] Step 2: Data preprocessing: After normalizing the collected multi-source signals, the sensors are used as nodes, and the cosine similarity between different sensors is calculated as edges to construct graph samples of different fault categories.

[0008] Step 3: In the training phase, the multi-layer graph isomorphism network model is trained using training samples from graph samples of different fault categories, and the parameters are continuously optimized;

[0009] Step 4: In the testing phase, the test samples from the different fault category graph samples are input into the trained multi-layer graph isomorphism network model, the probability distribution of the sample faults is output, and the sample label with the fault category with the highest probability is obtained as the diagnosis;

[0010] Step 5: Output the diagnosis results.

[0011] Furthermore, in step 2, the method for normalizing the collected multi-source signals is Min-Max normalization to obtain motor fault signal data.

[0012] Furthermore, in step 2, the steps of constructing graph samples of different fault categories are as follows:

[0013] In a graph sample, each sensor is mapped to a node in the graph. The initial features of the node are composed of the time domain signal data collected by the sensor in a fixed period. Assuming that there are N sensors deployed in the motor and each sensor collects M data points in time period T, the feature vector of the i-th sensor can be expressed as:

[0014]

[0015] in, is the sampling value of the i-th sensor at the k-th time point;

[0016] In order to characterize the correlation between sensors, cosine similarity is used as the metric of edge weight; for any two sensors i and j, their edge weight The calculation formula is as follows:

[0017]

[0018] Among them, the elements represents the edge weight between sensors i and j. The numerator on the right side of the equation represents the dot product of the two feature vectors, and the denominator is the product of the modulo lengths of the two vectors.

[0019] The cosine similarity value range is [−1, 1]. The closer the value is to 1, the more consistent the signal change trends of the two sensors are; the closer the value is to -1, the opposite trend is; and the value of 0 indicates that there is no significant linear correlation between the two.

[0020] Based on the above definitions of nodes and edges, construct the graph structure:

[0021] Where V is the node set, representing all sensors, E is the edge set, representing the connection relationship between sensors, X is the node feature matrix, each row corresponds to the feature vector of a sensor, and W is the edge weight matrix;

[0022] In this way, multi-source signals are converted into graph-structured data, which can effectively capture the correlation between sensors and provide input for subsequent graph isomorphic networks.

[0023] Furthermore, in step 3, during the training phase, the multi-layer graph isomorphic network model is trained using training samples from graph samples of different fault categories. The specific process includes the following steps:

[0024] The graph sample G is input into the cascade structure of the multi-layer graph isomorphism network to extract the complex relationship between sensors. For the graph sample G, the first The layer node feature update formula is as follows:

[0025]

[0026] in, Representation node In the Representation after the layer, initial state is the raw data point sequence feature of the sensor, is a learnable or fixed scalar, is a multilayer perceptron that learns nonlinear features from data from a sensor and its neighborhood. Representation node In the The layer representation, For nodes The set of neighbor nodes of

[0027] The cascaded structure of multi-layer graph isomorphism networks can capture a wider range of sensor-related information layer by layer. Each layer of GIN aggregates features from sensors in the immediate neighborhood and gradually expands to more distant neighborhoods, thereby constructing a deeper sensor representation. This structure enables the model to learn rich spatiotemporal feature relationships from sensor networks.

[0028] The original feature map representation of the image sample is obtained by the following formula:

[0029]

[0030] in, It is a simple summation operation that ensures that the graph representation is invariant to the sensor order, i.e., permutation, and can effectively capture the global structural characteristics of the entire sensor network. Compared with other GNN architectures, the design of GIN enhances the robustness and expressiveness of the model to sensor data by using a simple summation aggregation operation instead of complex normalization or pooling operations, and obtains the final generated graph representation through the cascade structure of multi-layer graph isomorphic networks. It can better capture the global information of the entire sensor network;

[0031] Get the graph representation of each sample Finally, the relational feature extractor composed of a multi-head self-attention mechanism and a feedforward network extracts the relational features between samples across graphs and learns the similarities or differences between different samples.

[0032] The multi-head self-attention mechanism is the key to the relational feature extractor. Its core idea is to focus on different relational features through multiple attention heads to calculate the relationship between graphs and graph samples, as follows:

[0033]

[0034] in, The formula for calculating the attention mechanism is: is a normalization function used to generate attention output, Q, K, V represent the query, key and value matrices extracted from the graph representation, respectively, and are derived from the input token T to the d-dimensional space through linear projection. is the dimension of the key vector;

[0035] By calculating the dot product of the query and the key, then performing softmax normalization on the result and multiplying it by the value matrix, a weighted representation is obtained that highlights the most relevant relationships. This mechanism can assign higher weights to important features and suppress the impact of noise;

[0036] The unique relational features captured by each attention head in the multi-head self-attention mechanism are concatenated together and linearly converted into a comprehensive relational feature encoding:

[0037]

[0038] in, is the formula for multi-head self-attention, which represents the combination of multiple attention heads. Indicates concatenating the outputs of all attention heads. 、 、 are the outputs from the 1st, 2nd, and hth attention heads, respectively. is the output projection matrix;

[0039] The feature representation of the relationship between graph samples at the lth relationship layer can be expressed as:

[0040]

[0041] in represents the refined relational feature representation, The proposed relational feature extractor is used to extract the relational features between different graph samples. Sample diagram In all sample sets The original graph representation in The union of

[0042] Furthermore, in step 3, the parameters are continuously optimized, and the specific process includes the following steps:

[0043] A new dynamic feedback training strategy is designed as a key component for continuously optimizing model parameters. In order to minimize the difference between the original graph representation space and the relationship feature space, two classifiers are used to train the model. and Constructing a feedback loop mechanism, the relational features from the relational feature extractor are extracted and transmitted to the cascade structure of the multi-layer graph isomorphism network to establish a dynamic feedback loop. The loss function is as follows:

[0044]

[0045] in, Is from and Cross entropy loss for labeled graph samples, is the coefficient of the cross entropy loss function, is the coefficient of the divergence loss function;

[0046] Is a Kullback-Leibler divergence loss used to measure the original graph representation and relation-aware graph representation The difference between the predicted probability distributions is as follows:

[0047]

[0048] Among them, among them, is the Kullback-Leibler divergence, , yes The predicted probability distribution, , yes The predicted probability distribution, is the temperature parameter of distillation;

[0049] It is a feature similarity loss function, which is represented by reducing the original graph and relation-aware representation The L2 distance between them makes the two embeddings more similar, guiding the cascade structure of the multi-layer graph isomorphism network to understand the feature representation of the relational feature extractor. The formula is as follows:

[0050]

[0051] By designing a loss function of a dynamic feedback loop combined with relational feature extraction to continuously optimize parameters of the cascade structure of the multi-layer graph isomorphic network, the ability of the method proposed in the present invention to understand and utilize the relationship between samples of different fault category graphs can be significantly enhanced.

[0052] The present invention proposes a motor fault diagnosis method based on graph isomorphism networks and cross-graph relationship learning. The cascade structure of a multi-layer graph isomorphism network is used to extract the original graph feature representation from the graph sample composed of multiple source signals of the motor, and the relationship features between samples of different fault categories are learned through the relationship feature extractor. After multiple iterations of the new training strategy, motor fault diagnosis is achieved and better diagnosis results are obtained. The present invention provides a new idea for motor fault diagnosis. This method emphasizes learning the relationship features of different fault categories from graph samples composed of multiple time domain signals collected by different sensors, and uses a multi-head self-attention mechanism and a feedforward network to capture the complex relationship between input features. It realizes timely adjustment through a dynamic feedback training strategy, and can more effectively handle noise. In fault diagnosis applications, the present invention shows stronger robustness to noise interference, significantly improves the recognition accuracy of different fault categories in a noisy environment, and is therefore more suitable for fault diagnosis scenarios under strong noise conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the motor fault diagnosis method of the present invention;

[0054] Figure 2 This is a radar chart showing the accuracy results of ten experiments using different fault diagnosis methods after adding -10dB noise to the original signal.

[0055] Figure 3 This is a radar chart showing the accuracy results of ten experiments using different fault diagnosis methods after adding -5dB noise to the original signal.

[0056] Figure 4 This is a radar chart showing the accuracy results of ten experiments using different fault diagnosis methods after adding 0dB noise to the original signal.

[0057] Figure 5 This is a radar chart showing the accuracy results of ten experiments using different fault diagnosis methods after adding 5dB noise to the original signal.

[0058] Figure 6 This is a radar chart showing the accuracy results of ten experiments using different fault diagnosis methods after adding 10dB noise to the original signal. DETAILED DESCRIPTION

[0059] In order to enable the present invention to construct graph samples from multi-source signals, graph samples containing different health states are used as inputs of the model, and after training the model, an optimized multi-layer graph isomorphic network cascade structure is obtained. The purpose of using the trained multi-layer graph isomorphic network model to diagnose the samples to be tested is achieved. The present invention adopts the following embodiments.

[0060] like Figure 1As shown in FIG, a motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning has the following steps:

[0061] Step 1: Obtain multi-source signals of the motor under different health conditions and establish a set of original signal samples;

[0062] Step 2: Data preprocessing: After normalizing the collected multi-source signals, the sensors are used as nodes, and the cosine similarity between different sensors is calculated as edges to construct graph samples of different fault categories.

[0063] For M data points collected by N sensors deployed in the motor within a period T, the feature vector of the i-th sensor can be expressed as:

[0064]

[0065] in, is the sampling value of the i-th sensor at the k-th time point;

[0066] In a preferred embodiment, seven sensors are used, and the raw data from each sensor in each sample includes 1024 data points;

[0067] In order to characterize the correlation between sensors, the present invention adopts cosine similarity as the metric of edge weight; for any two sensors i and j, their edge weight The calculation formula is as follows:

[0068]

[0069] Among them, the elements represents the edge weight between sensors i and j. The numerator on the right side of the equation represents the dot product of the two feature vectors, and the denominator is the product of the modulo lengths of the two vectors.

[0070] Based on the above definitions of nodes and edges, construct the graph structure:

[0071]

[0072] Where V is the node set, representing all sensors, E is the edge set, representing the connection relationship between sensors, X is the node feature matrix, each row corresponds to the feature vector of a sensor, and W is the edge weight matrix;

[0073] In this way, multi-source signals are converted into graph-structured data, providing input for the subsequent cascade structure of multi-layer graph isomorphism networks;

[0074] Step 3: In the training phase, the multi-layer graph isomorphism network model is trained using training samples from graph samples of different fault categories, and the parameters are continuously optimized;

[0075] The graph sample G is input into the cascade structure of a multi-layer graph isomorphism network to extract the complex relationship between sensors. The process of graph isomorphism network feature extraction is as follows:

[0076] For the graph sample G, the graph isomorphism network The layer node feature update formula is as follows:

[0077]

[0078] in, Representation node In the Representation after the layer, initial state is the raw data point sequence feature of the sensor, is a learnable or fixed scalar, is a multilayer perceptron that learns nonlinear features from data from a sensor and its neighborhood. Representation node In the The layer representation, For nodes The set of neighbor nodes of

[0079] The original feature map representation of the image sample is obtained by the following formula:

[0080]

[0081] Get the graph representation of each sample Finally, the relational features between different samples are learned through a relational feature extractor composed of a multi-head self-attention mechanism and a feedforward network:

[0082]

[0083] in, The formula for calculating the attention mechanism is: is a normalization function used to generate attention output, Q, K, V represent the query, key and value matrices extracted from the graph representation, respectively, and are derived from the input token T to the d-dimensional space through linear projection. is the dimension of the key vector, which represents the structure and features of the encoded graph;

[0084] The relational features captured by each head in the multi-head self-attention mechanism are spliced ​​together and first converted into a comprehensive relational feature encoding:

[0085]

[0086] in, is the formula for multi-head self-attention, which represents the combination of multiple attention heads. Indicates concatenating the outputs of all attention heads. 、 、 are the outputs from the 1st, 2nd, and hth attention heads, respectively. is the output projection matrix;

[0087] The feature representation of the relationship between graph samples at the lth relationship layer can be expressed as:

[0088]

[0089] in, represents the refined relational feature representation, The proposed relational feature extractor is used to extract the relational features between different graph samples. Sample diagram In all sample sets The original graph representation in The union of

[0090] Through two classifiers and Constructing a feedback loop mechanism, the relational features from the relational feature extractor are extracted and transmitted to the cascade structure of the multi-layer graph isomorphism network to establish a dynamic feedback loop. The loss function is as follows:

[0091]

[0092] in, Is from and Cross entropy loss for labeled graph samples, is the coefficient of the cross entropy loss function, is the coefficient of the divergence loss function;

[0093] Is a Kullback-Leibler divergence loss used to measure the original graph representation and contains the relationship feature graph representation The difference between the predicted probability distributions is as follows:

[0094]

[0095] Among them, among them, is the Kullback-Leibler divergence, , yes The predicted probability distribution, , yes The predicted probability distribution, is the temperature parameter of distillation;

[0096] It is a feature similarity loss function, which is represented by reducing the original graph and relation-aware representation The L2 distance between them makes the two embeddings more similar, guiding the cascade structure of the multi-layer graph isomorphism network to understand the feature representation of the relational feature extractor. The formula is as follows:

[0097]

[0098] The loss function of the designed dynamic feedback loop combined with relational feature extraction is used to continuously optimize the parameters of the cascade structure of the multi-layer graph isomorphism network.

[0099] Step 4: In the testing phase, the test samples from the different fault category graph samples are input into the trained multi-layer graph isomorphism network model, and the probability distribution of the sample faults is output. The fault category with the highest probability is obtained as the predicted sample label;

[0100] Step 5: Output the diagnosis results.

[0101] The present invention is verified by the following experiments.

[0102] To verify the effectiveness of the proposed method, the operating conditions and health status of the test sample were pre-set during the experiment, and five different conditions of the motor under test were established for analysis, including the normal state of the motor, three bearing fault conditions (inner ring fault, outer ring fault, and outer ring combined fault), and a magnet crack fault condition.

[0103] To obtain multi-source signals from the motor under test, a total of seven signals were collected: three vibration signals, one sound pressure signal, and three current signals. A three-axis vibration accelerometer (x, y, and z) was mounted on the motor housing, and a microphone was placed 30 cm from the end cap of the test motor. During the experiment, the motor under test rotated at 2800 RPM and the sampling frequency was 48,000 Hz. In the following analysis, samples under different conditions included one sound pressure signal, three acceleration vibration signals, and three current signals. Each signal in each sample had 1024 consecutive data points.

[0104] To verify the effectiveness of the proposed method in motor fault diagnosis, a comparative analysis was conducted using graph neural network models: GCN, GIN, ChebyNet, GraphSage, GAT, HoGCN, and SGCN. To evaluate MsEGIN's noise immunity, Gaussian white noise with varying signal-to-noise ratios (SNRs) of 10dB, 5dB, 0dB, −5dB, and −10dB was added to the original signal. Each state (normal, bearing inner race fault, outer race fault, magnet fault, and combined bearing inner and outer race failure) consisted of 200 samples, 120 of which were used for training and 80 for testing, resulting in a total of 400 test samples. The experiment was repeated ten times, and the average results of these ten iterations are reported.

[0105] In motor fault diagnosis applications, diagnostic accuracy (Acc), F1-score (F1-score), precision (Precision), and recall (Recall) are of particular interest. Therefore, these four metrics are used to quantitatively compare the diagnostic effectiveness of different methods.

[0106] Table 1 compares the evaluation indicators of different methods

[0107] method Avg-Acc F1-score Precision Recall GCN 97.00 97.65 97.79 97.65 GIN 98.4 98.39 98.46 98.40 ChebyNet 97.80 97.80 97.93 97.80 GraphSage 91.10 89.49 90.49 91.10 GAT 97.85 97.85 98.02 97.85 HoGCN 86.80 84.76 89.96 86.80 SGCN 81.40 77.15 80.91 81.40 Method of the present invention 99.58 99.60 99.64 99.60

[0108] Table 1

[0109] The experiment used multiple graph neural network models to conduct comparative analysis with the method proposed in the present invention. From the results in Table 1, it can be seen that the method proposed in the present invention exhibits the best performance. Compared with the second best method GIN in the table, it achieved an average accuracy improvement of 1.18%, an F1 score improvement of 1.21%, an accuracy improvement of 1.18%, and a recall improvement of 1.2%. The method of the present invention performed well in multiple key indicators and achieved a significant improvement in average accuracy, reflecting the model's good classification ability for different fault categories. The outstanding performance in precision and recall shows the superiority of the method of the present invention in reducing false positives and missed faults, and it can more effectively capture faults of different categories than other models.

[0110] To evaluate the noise immunity of our method, we added Gaussian white noise with varying signal-to-noise ratios (SNRs) to the original signal to test its robustness. The average accuracy of the different methods across ten experiments under varying noise conditions is shown in Table 2.

[0111] Table 2 shows the average accuracy (%) of ten experiments for different methods at different signal-to-noise ratios:

[0112] method -10db -5db 0db 5db 10db GCN 37.25 47.45 62.05 72.80 86.70 GIN 50.35 59.15 70.58 83.75 94.95 ChebyNet 36.9 47.7 62.15 71.55 86.25 GraphSage 39.2 48.2 65.4 72.1 91.85 GAT 36.65 46.75 64.85 72.45 85.55 HOGCN 40.6 48.6 65.3 72.45 88.50 SGCN 37.75 46.85 68.9 74.85 86.10 Method of the present invention 79.30 81.65 90.30 91.80 96.35

[0113] Table 2

[0114] As shown in Table 2, as the signal-to-noise ratio decreases, the performance of each model decreases. Due to different robustness to noise, the accuracy drop rates of these methods are also different. Compared with other methods, even at -10dB SNR, the method of the present invention can still obtain the best diagnosis results, which proves that the noise resistance of the method of the present invention is better than that of the comparison method. The radar chart of the accuracy of ten experiments of different methods under different noise conditions is shown in Figure 2-Figure 6 .from Figure 2-Figure 6 As can be seen, the proposed method consistently outperforms other GNN models across all noise levels. Under extreme noise conditions (SNR = -10), the accuracy of other methods drops sharply, while the proposed method performs relatively stably and exhibits better noise resistance. This is due to the multi-head self-attention mechanism in the relational feature extractor, which dynamically assigns attention weights based on the importance of input features. This allows the proposed method to focus on important features while ignoring less important noise features. Consequently, the proposed method exhibits greater robustness in the presence of noisy data. Experimental results demonstrate that the proposed method can accurately identify motor faults under varying noise levels compared to other methods, demonstrating its superior stability and noise resistance.

Claims

1. A motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning, characterized in that: The following steps are involved: Step 1: Obtain multi-source signals of the motor in different health states and establish a set of original signal samples; Step 2: Data preprocessing: After normalizing the collected multi-source signals, the sensors are used as nodes, and the cosine similarity between different sensors is calculated as edges to construct graph samples of different fault categories. Step 3: In the training phase, the multi-layer graph isomorphism network model is trained using training samples from graph samples of different fault categories, and the parameters are continuously optimized; Step 4: In the testing phase, the test samples from the different fault category graph samples are input into the trained multi-layer graph isomorphism network model, and the probability distribution of the sample faults is output. The fault category with the highest probability is obtained as the predicted sample label; Step 5: Output the diagnosis results; The steps to construct graph samples of different fault categories are as follows: In a graph sample, each sensor is mapped to a node in the graph. The initial features of the node are composed of the time domain signal data collected by the sensor in a fixed period. Assuming that there are N sensors deployed in the motor and each sensor collects M data points in time period T, the feature vector of the i-th sensor can be expressed as: ； in, is the sampling value of the i-th sensor at the k-th time point; In order to characterize the correlation between sensors, cosine similarity is used as the metric of edge weight; For any two sensors i and j, their edge weights The calculation formula is as follows: ； Among them, the elements represents the edge weight between sensors i and j. The numerator on the right side of the equation represents the dot product of the two feature vectors, and the denominator is the product of the modulo lengths of the two vectors. Based on the above definitions of nodes and edges, we construct graph samples of different fault categories: ； Where V is the node set, representing all sensors, E is the edge set, representing the connection relationship between sensors, X is the node feature matrix, each row corresponds to the feature vector of a sensor, and W is the edge weight matrix; In step 3, the parameters are continuously optimized, and the specific process includes the following steps: Through two classifiers and Constructing a feedback loop mechanism, the relational features from the relational feature extractor are extracted and transmitted to the cascade structure of the multi-layer graph isomorphism network to establish a dynamic feedback loop. The loss function is as follows: ； in, Is from and Cross entropy loss for labeled graph samples, is the coefficient of the cross entropy loss function, is the coefficient of the divergence loss function; Is a Kullback-Leibler divergence loss used to measure the original graph representation and contains the relationship feature graph representation The difference between the predicted probability distributions is as follows: ； in, is the Kullback-Leibler divergence, , yes The predicted probability distribution, , yes The predicted probability distribution, is the temperature parameter of distillation; It is a feature similarity loss function, which is represented by reducing the original graph and relation-aware representation The L2 distance between them makes the two embeddings more similar, guiding the cascade structure of the multi-layer graph isomorphism network to understand the feature representation of the relational feature extractor. The formula is as follows: ； The loss function of the designed dynamic feedback loop combined with relational feature extraction is used to continuously optimize the parameters of the cascade structure of the multi-layer graph isomorphism network.

2. A motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning according to claim 1, characterized in that: In step 2, the method for normalizing the collected multi-source signals is Min-Max normalization to obtain motor fault signal data.

3. The motor fault diagnosis method based on graph isomorphism network and cross-graph relationship learning according to claim 1 is characterized in that: In step 3, during the training phase, the multi-layer graph isomorphic network model is trained using training samples from graph samples of different fault categories. The specific process includes the following steps: The graph sample G is input into the cascade structure of the multi-layer graph isomorphism network to extract the complex relationship between sensors. For the graph sample G, the first The layer node feature update formula is as follows: ； in, Representation node In the Representation after the layer, initial state is the raw data point sequence feature of the sensor, is a learnable or fixed scalar, is a multilayer perceptron that learns nonlinear features from data from a sensor and its neighborhood. Representation node In the The layer representation, For nodes The set of neighbor nodes of The original feature map representation of the image sample is obtained by the following formula: ； in, It is a simple sum operation; Get the graph representation of each sample Finally, the relational features between different samples are learned through a relational feature extractor composed of a multi-head self-attention mechanism and a feedforward network: ； in, The formula for calculating the attention mechanism is: is a normalization function used to generate attention output, Q, K, V represent the query, key and value matrices extracted from the graph representation, respectively, and are derived from the input token T to the d-dimensional space through linear projection. is the dimension of the key vector; The relational features captured by each head in the multi-head self-attention mechanism are spliced ​​together and first converted into a comprehensive relational feature encoding: ； in, is the formula for multi-head self-attention, which represents the combination of multiple attention heads. Indicates concatenating the outputs of all attention heads. 、 、 are the outputs from the 1st, 2nd, and hth attention heads, respectively. is the output projection matrix; The feature representation of the relationship between graph samples at the lth relationship layer is expressed as: ； in, represents the refined relational feature representation, The proposed relational feature extractor is used to extract the relational features between different graph samples. Sample diagram In all sample sets The original graph representation in The union of .

Images