A planetary gearbox fault diagnosis method based on physical information enhanced graph attention network

Abstract

The present application relates to the technical field of fault diagnosis, and particularly relates to a planetary gearbox fault diagnosis method based on a physical information enhanced graph attention network. The method extracts physical features and statistical features, constructs a hypergraph structure through set operation and comprehensive weighting, solves the problem of ignoring physical information in a single feature space, introduces physical prior edge weight to guide attention calculation through a feature coding fusion module and a physically guided graph attention mechanism, and mines fault mechanism features in a vibration signal. The method realizes probability level fusion of physical knowledge and data driving by constraining attention distribution approximation to physical prior through a KL divergence regular term. The present application can effectively solve problems such as poor graph quality of a graph neural network, missing physical information and the like under strong noise interference, and realizes efficient and accurate diagnosis of weak faults of a planetary gearbox.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a fault diagnosis method for planetary gearboxes based on a physical information augmented graph attention network. Background Technology

[0002] Planetary gearboxes, as core components of transmission systems in high-end equipment such as mining machinery, wind power generation, aerospace, and heavy machinery, possess advantages such as high power density, large transmission ratio, and compact structure. However, operating under harsh environments with heavy loads and high noise levels for extended periods, critical components such as the sun gear, planet gears, ring gear, and bearings are highly susceptible to failure, potentially leading to equipment downtime or even major safety accidents. Therefore, conducting research on planetary gearbox fault diagnosis is of great significance for ensuring safe equipment operation and reducing maintenance costs.

[0003] In recent years, with the development of the Industrial Internet of Things (IIoT) and the arrival of the big data era, equipment condition monitoring data has experienced explosive growth. Deep learning technology, with its powerful feature self-learning capabilities, has become a research hotspot in the field of fault diagnosis. Convolutional neural networks (CNNs), with their powerful local feature extraction capabilities, automatically learn fault features in vibration signals through multi-layer convolution and pooling operations; recurrent neural networks (RNNs) and their variant, long short-term memory (LSTM) networks, excel at processing time-series data and can capture the temporal dependencies of fault signals. However, these networks typically assume that samples are independent of each other, making it difficult to effectively utilize the correlation information between samples, and they are also sensitive to noise interference, limiting their diagnostic performance in noisy environments.

[0004] The rise of graph neural networks has provided a new technical path for fault diagnosis. Graph networks can handle non-Euclidean structured data and effectively mine the correlations between samples through message passing mechanisms between nodes. The excellent diagnostic performance of graph networks largely depends on the quality of graph data construction. Existing graph network-based fault diagnosis methods typically use the K-nearest neighbor algorithm, combining metrics such as Euclidean distance or cosine similarity between data waveforms to calculate the similarity between samples, and then selecting the K most similar neighbors for each sample to establish edge relationships. In addition, composite graph construction strategies often employ the fusion of multiple feature spaces or the use of multiple metrics, taking the intersection or union of graph structures to enrich the correlation information between samples, in order to obtain a more comprehensive graph structure representation.

[0005] However, existing graph construction methods still have the following limitations: 1) Whether it is a single graph or a composite graph, they only rely on the trend characteristics of the data waveform, ignoring the physical mechanism information contained in the samples. This makes it difficult for the constructed graph structure to truly reflect the intrinsic relationship between fault samples, limiting the model's generalization ability in a strong noise background. 2) Simple intersection and union operations in composite graph construction are difficult to effectively balance the complementarity and redundancy of multi-source information. In a strong noise environment, they are prone to introducing noise edges, interfering with model learning. 3) Most existing methods only utilize the spatial dimension features of the signal and fail to effectively integrate multi-dimensional information such as time domain and frequency domain, making it difficult to comprehensively characterize fault features. Summary of the Invention

[0006] To overcome the above shortcomings, this invention provides a planetary gearbox fault diagnosis method based on a physically-informed graph attention network. The aim is to integrate the fault mechanism of the planetary gearbox into the graph structure design, construct a hypergraph structure that can comprehensively characterize the sample association, and design a graph neural network model with physical guidance capabilities to achieve high-precision and high-robust fault diagnosis in a noisy environment.

[0007] This invention provides the following technical solution: a planetary gearbox fault diagnosis method based on a physically enhanced graph attention network, comprising: Step 1: Collect vibration acceleration signals of the planetary gearbox under different fault modes; Step 2: Standardize the vibration acceleration signal and divide the samples to construct a training dataset and a test dataset; Step 3: Perform physical information augmentation on the training and test datasets, extract physical feature vectors and statistical feature vectors, establish edge relationships based on the similarity between feature vectors, and construct training graph data and test graph data; Step 4: Construct a physical information-enhanced graph attention network, train the network parameters using training graph data, and constrain it using a hybrid loss function to obtain a fault diagnosis model; Step 5: Input the test diagram data into the fault diagnosis model to identify and output the fault mode of the planetary gearbox.

[0008] Preferably, the steps for constructing the training dataset and the test dataset include: The original vibration acceleration signal is standardized by using the mean function and the standard deviation function to obtain the standardized vibration signal; The standardized vibration signal is divided into multiple fixed-length sub-samples by setting a sliding window length and using a non-overlapping sliding window resampling method to form a sample set. Set the sample ratio between the training set and the test set, and randomly select non-overlapping sub-samples from the sample set to serve as the training set and the test set, respectively.

[0009] Preferably, the steps for extracting physical feature vectors and statistical feature vectors include: The spectra of the training dataset and the test dataset are calculated by Fourier transform, and spectral feature vectors are extracted from the spectra. The envelope spectra of the training dataset and the test dataset are calculated using Hilbert transform, and envelope feature vectors are extracted from the envelope spectra. The spectral feature vector and the envelope feature vector are concatenated to form a physical feature vector; Statistical features are extracted from data subsamples in the training dataset and the test dataset to form a statistical feature vector.

[0010] Preferably, the step of establishing edge relationships based on the similarity between feature vectors includes: The Euclidean distance between each sample is calculated using the K-nearest neighbor method for both the physical feature vector and the statistical feature vector. Select the k nearest neighbors and construct edge relationships based on physical feature vectors and edge relationships based on statistical feature vectors, respectively.

[0011] Preferably, the steps for constructing training graph data and test graph data include: The edge relationships based on physical feature vectors and the edge relationships based on statistical feature vectors are merged using a weighted fusion method to obtain a comprehensive edge relationship; Using sample data as node features, we construct training and test graph data with enhanced physical information.

[0012] Preferably, the physically-enhanced graph attention network includes a feature encoding fusion module, a physically-guided graph attention module, and a classification module; wherein, the feature encoding fusion module consists of a data splicing layer and two fully connected network layers, the physically-guided graph attention module consists of two graph attention network layers using edge relation weights in the training graph data or test graph data, and the classification module consists of two fully connected network layers.

[0013] Preferably, the computation process of the physically guided graph attention module includes: For each node, the attention coefficients between it and its neighboring nodes are calculated through a shared attention mechanism, and then a nonlinear transformation is performed using the LeakyReLU activation function; The transformation result is normalized using the softmax function to obtain normalized attention coefficients, which are then used for weighted aggregation of neighbor node features.

[0014] Preferably, the step of training the network parameters includes: initializing the network parameters, setting the optimizer, initial learning rate, maximum training steps and loss function, and iteratively training the network using the training graph data until the maximum training steps or loss function convergence is reached.

[0015] Preferably, the step of constraining using a hybrid loss function includes: Cross-entropy loss is used to measure the difference between the model's predicted category and the true label; The similarity between the distribution of attention coefficients in the KL divergence loss constrained graph attention network and the distribution of the integrated edge relation weights in the training graph data is calculated. The cross-entropy loss and the KL divergence loss are weighted and summed to form the final mixed loss function.

[0016] The present invention has the following beneficial effects: 1. In this invention, physical and statistical features are extracted based on the fault mechanism of planetary gearboxes. Physical information enhancement graph data is constructed through union operation and comprehensive weighting. The physical mechanism of the fault is integrated into the graph structure design, which effectively solves the problem of single feature space graph construction and ignoring physical information, and provides high-quality input for graph neural networks.

[0017] 2. This invention introduces physical prior edge weights to guide attention calculation, enabling the model to focus on sample relationships that conform to the fault mechanism during feature aggregation. At the same time, the feature encoding fusion module simultaneously utilizes spectral features and physical features, which significantly improves the model's feature extraction capability and diagnostic accuracy under strong noise interference.

[0018] 3. This invention combines the KL divergence regularization term with the cross-entropy loss to constrain the attention distribution to approximate the physical prior distribution, so that the model maintains consistency with physical laws during the learning process, effectively enhancing the training stability and generalization performance of the model. Attached Figure Description

[0019] Figure 1 This is a flowchart of a planetary gearbox fault diagnosis method based on a physically enhanced graph attention network proposed in this invention. Figure 2 This is a schematic diagram of a planetary gearbox fault simulation test bench in an embodiment of the present invention; In the diagram, 1 is the mounting bracket, 2 is the drive motor, 3 is the coupling I, 4 is the planetary gearbox, 5 is the coupling II, 6 is the magnetic powder brake, and 7 is the base plate. Figure 3 These are time-domain waveforms of vibration acceleration of key components of the planetary gearbox under different fault modes in embodiments of the present invention. Figure 4This is a schematic diagram of the physical information enhancement graph attention network structure of the present invention; Figure 5 This is the confusion matrix of diagnostic results in the embodiments of the present invention.

[0020] In the diagram, 0 is normal, 1 is sun gear wear, 2 is sun gear peeling, 3 is sun gear crack, 4 is sun gear broken tooth, 5 is sun gear missing tooth, 6 is planet gear peeling, 7 is planet gear crack, 8 is planet gear broken tooth, 9 is planet gear missing tooth, 10 is planet bearing rolling element, 11 is planet bearing outer ring, 12 is compound fault one: sun gear broken tooth + planet bearing inner ring, 13 is compound fault two: sun gear broken tooth + planet bearing outer ring. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The flowchart of the planetary gearbox fault diagnosis method based on physically enhanced graph attention network of this invention is as follows: Figure 1 As shown, the planetary gearbox failure simulation test bench used in the embodiment is as follows: Figure 2 As shown, the experimental platform consists of a mounting bracket 1, a drive motor 2, a coupling I3, a planetary gearbox 4, a coupling II5, a magnetic powder brake 6, and a base plate 7. In this embodiment, the drive motor speed is set to 150 rpm, and the magnetic powder brake provides a maximum load torque of 15 Nm. The vibration data acquisition system consists of an NI9231 data acquisition card, an industrial control computer with DASP software, and an accelerometer of model Brüel & Kjær 8344. The sampling frequency is set to 12800 Hz, and the sampling duration is set to 300 s. The basic parameters of the key components of the planetary gearbox are shown in Table 1.

[0023] Table 1 Basic parameters of key components of planetary gearbox

[0024] This invention provides a planetary gearbox fault diagnosis method based on a physically enhanced graph attention network, the specific steps of which are as follows: Step 1: Collect vibration acceleration signals of the planetary gearbox under different fault modes; Specifically, vibration acceleration signals of the planetary gearbox under different failure modes of its key components were collected. Where i = 1, 2, ..., Z, and Z is the total number of failure modes. Vibration acceleration signals obtained under different fault modes The total number of sampling points included; in this embodiment, the key components of the planetary gearbox include the sun gear, planet gears, ring gear, and planetary bearings. There are 14 fault modes, and the time-domain waveform of the vibration acceleration signal collected under each fault mode is as follows: Figure 3 As shown.

[0025] Step 2: Standardize the vibration acceleration signal and divide the samples to construct a training dataset and a test dataset; Preferably, the steps for constructing the training dataset and the test dataset include: The original vibration acceleration signal is standardized by using the mean function and the standard deviation function to obtain the standardized vibration signal; The standardized vibration signal is divided into multiple fixed-length sub-samples by setting a sliding window length and using a non-overlapping sliding window resampling method to form a sample set. Set the sample ratio between the training set and the test set, and randomly select non-overlapping sub-samples from the sample set to serve as the training set and the test set, respectively.

[0026] Specifically, the vibration signal Standardization processing is performed to obtain a standardized vibration acceleration signal. ,in: ; Where Mean(·) is the mean function and Std(·) is the standard deviation function; (2) Set the length of the sliding window M=12800, and use the non-overlapping sliding window resampling method to resample the standardized signal in step (1). In one feasible implementation, the signal is divided into standardized signals. The dataset is divided into 4200 subsamples of length 12800, and the subsamples are used to construct dataset D. ; in, floor(·) is the floor function; (3) Set the ratio of the number of samples in the training set and the test set, and randomly select the corresponding ratio of data from the dataset D as the training set. and test set Furthermore, there are no duplicate samples in the training set and the test set; in one feasible implementation, the sample ratio of the training set and the test set is 4:1, and the detailed information of the split dataset is shown in Table 2.

[0027] Table 2 Dataset Details

[0028] Step 3: Perform physical information augmentation on the training and test datasets, extract physical feature vectors and statistical feature vectors, establish edge relationships based on the similarity between feature vectors, and construct training graph data and test graph data; Preferably, the steps for extracting physical feature vectors and statistical feature vectors include: The spectra of the training dataset and the test dataset are calculated by Fourier transform, and spectral feature vectors are extracted from the spectra. The envelope spectra of the training dataset and the test dataset are calculated using Hilbert transform, and envelope feature vectors are extracted from the envelope spectra. The spectral feature vector and the envelope feature vector are concatenated to form the physical feature vector. Statistical features are extracted from data subsamples in the training dataset and the test dataset to form a statistical feature vector.

[0029] Specifically, the spectra of data subsamples in the training and test sets are calculated. ; in, For the training set spectrum, For the spectrum of the training set, FFT(·) denotes Fourier transform; Extract the meshing frequency from the obtained spectrum. The spectral characteristic vectors in the sidebands include meshing harmonic energy, meshing harmonic peak value, planetary gear passing frequency sideband energy, sun gear distributed fault sideband energy, planetary gear distributed fault sideband energy, and ring gear distributed fault sideband energy, as shown in the following formula: ; ; ; ; ; ; Where i is the meshing harmonic order, j is the sideband modulation order of the planetary gear passing frequency, and k is the sideband modulation order of the distributed fault. , , , , , These represent the energy of the i-th order meshing harmonic, the peak value of the meshing harmonic, the sideband energy of the planetary gear passing frequency, the sideband energy of the sun gear distributed fault, the sideband energy of the planetary gear distributed fault, and the sideband energy of the ring gear distributed fault, respectively. , , , , These are, respectively, the meshing frequency, the planetary gear passing frequency, the sun gear distributed rotational frequency, the planetary gear distributed rotational frequency, and the ring gear frequency. This is the spectrum function.

[0030] Then, the envelope spectra of data subsamples in the training and test sets are calculated. ; in, To train the envelope spectrum, For the test set envelope spectrum, H(·) denotes the Hilbert transform. From the obtained envelope spectrum, the envelope feature vectors of key components of the planetary gearbox are extracted, including the local fault feature energy of the sun gear, planet gears, ring gear, and bearings, as shown in the following equation: ; ; ; ; in, , , , Let represent the characteristic energies of local faults in the i-th order sun gear, planet gears, ring gear, and bearing, respectively. , , , These are the characteristic frequencies of failures in the inner ring, outer ring, rolling elements, and cage of the planetary bearing, respectively. The energy spectrum is then used. The spectral eigenvector and envelope eigenvector are subsequently concatenated into a single physical eigenvector. .

[0031] Finally, the statistical characteristics of the data subsamples in the training and test sets are calculated, including mean, standard deviation, root mean square value, skewness, kurtosis, kurtosis, peak factor, and impulse factor, and the statistical characteristics of each subsample are constructed into a statistical feature vector. .

[0032] Preferably, the step of establishing edge relationships based on the similarity between feature vectors includes: The Euclidean distance between each sample is calculated using the K-nearest neighbor method for both the physical feature vector and the statistical feature vector. Select the k nearest neighbors and construct edge relationships based on physical feature vectors and edge relationships based on statistical feature vectors, respectively.

[0033] Specifically, the K-nearest neighbor method is used to calculate the physical feature vector corresponding to each subsample. The Euclidean distance between them is used to construct the edge relationship by selecting the k nearest neighbors. The node features are the original subsample data in the training set and the test set. ; Where KNN(.) represents the K-nearest neighbor function, k represents the number of nearest neighbors (k=5), and N Let be the set of neighbors of node i calculated based on its physical feature vector. Let be the physical feature vector of node i. This represents the Euclidean distance between node i and node j, calculated based on their physical feature vectors. The edge relationship weight between node i and node j is calculated based on the physical feature vector; Similarly, for statistical feature vectors The K-nearest neighbor method is used to establish the edge relationships between each subsample, as shown in the following formula: ; in, Let i be the set of neighbors calculated based on the statistical feature vector. Let be the statistical feature vector of node i. This represents the Euclidean distance between node i and node j, calculated based on their statistical eigenvectors. The edge relationship weight between node i and node j is calculated based on the statistical feature vector.

[0034] Preferably, the steps for constructing training graph data and test graph data include: The edge relationships based on physical feature vectors and the edge relationships based on statistical feature vectors are merged using a weighted fusion method to obtain a comprehensive edge relationship; Using sample data as node features, we construct training and test graph data with enhanced physical information.

[0035] Specifically, a union operation is used to fuse the calculated edge relationships between nodes to obtain a training set of physical information augmented graph data. Physical information augmented graph data test set The edge relationship weights are described by the comprehensive similarity, as shown in the following formula: ; in, , As an adjustable weighting parameter, in one feasible implementation, =0.8, =0.2.

[0036] Step 4: Construct a physical information-enhanced graph attention network, train the network parameters using training graph data, and constrain it using a hybrid loss function to obtain a fault diagnosis model; Preferably, the physically-enhanced graph attention network includes a feature encoding fusion module, a physically-guided graph attention module, and a classification module; wherein, the feature encoding fusion module consists of a data splicing layer and two fully connected network layers, the physically-guided graph attention module consists of two graph attention network layers using edge relation weights in the training graph data or test graph data, and the classification module consists of two fully connected network layers.

[0037] Preferably, the computation process of the physically guided graph attention module includes: For each node, the attention coefficients between it and its neighboring nodes are calculated through a shared attention mechanism, and then a nonlinear transformation is performed using the LeakyReLU activation function; The transformation result is normalized using the softmax function to obtain normalized attention coefficients, which are then used for weighted aggregation of neighbor node features.

[0038] Preferably, the step of training the network parameters includes: initializing the network parameters, setting the optimizer, initial learning rate, maximum training steps and loss function, and iteratively training the network using the training graph data until the maximum training steps or loss function convergence is reached.

[0039] Specifically, a physical information-enhanced graph attention network is constructed, the framework of which is shown in the figure below. Figure 4 As shown. The first part of the network is the feature encoding fusion module, the second part is the physically guided graph attention module, and the third part is the classification module. The feature encoding fusion module consists of data concatenation and two fully connected layers. The physically guided graph attention module consists of two layers that use physical information to enhance the edge relationship weights in the graph data. The graph attention network consists of the following components, and the computation process of the physically guided graph attention module is shown in the following formula: ; ; Where W is the learnable weight matrix, and Let i and j be the node features of nodes i and j, respectively, and LeakyReLU be the activation function. For data splicing symbols, a shared attention mechanism It is a single-layer feedforward neural network. Let i be the set of neighbors of node i.

[0040] The network parameters were then initialized, and the physical information augmentation graph data training set was used. The input is fed into a Physical Information Augmentation Graph Attention Network for feature extraction. In one feasible implementation, the Xavier initialization method is used to initialize all learnable weight matrices and bias terms in the network; the optimizer is set to the Adam optimizer with an initial learning rate of 0.001, a maximum training step of 500 epochs, and an early stopping strategy is adopted (if the validation set loss does not decrease for 20 consecutive epochs, the network is terminated early).

[0041] Preferably, the step of constraining using a hybrid loss function includes: Cross-entropy loss is used to measure the difference between the model's predicted category and the true label; The similarity between the distribution of attention coefficients in the KL divergence loss constrained graph attention network and the distribution of the integrated edge relation weights in the training graph data is calculated. The cross-entropy loss and the KL divergence loss are weighted and summed to form the final mixed loss function.

[0042] Specifically, based on the data sample labels and the edge relationship weights in the physical information augmented graph data... The network parameters are trained using a hybrid loss function with dual constraints of cross-entropy and KL divergence. The hybrid loss function is shown in the following equation:

[0043] in, Let c be the c-th value in the one-hot encoding corresponding to the true label of the i-th sample. To predict the probability that the i-th sample belongs to class c, where C is the total number of classes. The attention coefficient is the edge relationship between nodes i and j. The neighboring nodes of node i. Based on the set maximum number of training steps, repeat step 4 to obtain the trained physical information augmented graph attention network. Step 5: Input the test diagram data into the fault diagnosis model to identify and output the fault modes of the planetary gearbox. In one feasible implementation, the confusion matrix of the diagnostic results is as follows: Figure 5 As shown, only one planetary bearing rolling element fault sample was misdiagnosed and incorrectly classified as normal, while the remaining samples were correctly classified, achieving a classification accuracy of 99.9%. The diagnostic results indicate that the planetary gearbox fault diagnosis method based on a physical information-enhanced graph attention network provided by this invention can effectively solve problems such as poor graph quality and missing physical information under strong noise interference, and achieve efficient and accurate diagnosis of minor faults in planetary gearboxes.

[0044] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A planetary gearbox fault diagnosis method based on a physically-informed graph attention network, characterized in that, include: Step 1: Collect vibration acceleration signals of the planetary gearbox under different fault modes; Step 2: Standardize the vibration acceleration signal and divide the samples to construct a training dataset and a test dataset; Step 3: Perform physical information augmentation on the training and test datasets, extract physical feature vectors and statistical feature vectors, establish edge relationships based on the similarity between feature vectors, and construct training graph data and test graph data; Step 4: Construct a physical information-enhanced graph attention network, train the network parameters using training graph data, and constrain it using a hybrid loss function to obtain a fault diagnosis model; Step 5: Input the test diagram data into the fault diagnosis model to identify and output the fault mode of the planetary gearbox.

2. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 1, characterized in that, The steps to build the training and testing datasets include: The original vibration acceleration signal is standardized by using the mean function and the standard deviation function to obtain the standardized vibration signal; The standardized vibration signal is divided into multiple fixed-length sub-samples by setting a sliding window length and using a non-overlapping sliding window resampling method to form a sample set. Set the sample ratio between the training set and the test set, and randomly select non-overlapping sub-samples from the sample set to serve as the training set and the test set, respectively.

3. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 1, characterized in that, The steps for extracting physical feature vectors and statistical feature vectors include: The spectra of the training dataset and the test dataset are calculated by Fourier transform, and spectral feature vectors are extracted from the spectra. The envelope spectra of the training dataset and the test dataset are calculated using Hilbert transform, and envelope feature vectors are extracted from the envelope spectra. The spectral feature vector and the envelope feature vector are concatenated to form the physical feature vector. Statistical features are extracted from data subsamples in the training dataset and the test dataset to form a statistical feature vector.

4. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 3, characterized in that, The steps for establishing edge relationships based on the similarity between feature vectors include: The Euclidean distance between each sample is calculated using the K-nearest neighbor method for both the physical feature vector and the statistical feature vector. Select the k nearest neighbors and construct edge relationships based on physical feature vectors and edge relationships based on statistical feature vectors, respectively.

5. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 4, characterized in that, The steps for constructing training and testing graph data include: The edge relationships based on physical feature vectors and the edge relationships based on statistical feature vectors are merged using a weighted fusion method to obtain a comprehensive edge relationship; Using sample data as node features, we construct training and test graph data with enhanced physical information.

6. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 1, characterized in that, The physical information-enhanced graph attention network includes a feature encoding fusion module, a physical-guided graph attention module, and a classification module; wherein, the feature encoding fusion module consists of a data splicing layer and two fully connected network layers, the physical-guided graph attention module consists of two graph attention network layers that use the edge relation weights in the training graph data or test graph data, and the classification module consists of two fully connected network layers.

7. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 6, characterized in that, The computation process of the physically guided graph attention module includes: For each node, the attention coefficients between it and its neighboring nodes are calculated through a shared attention mechanism, and then a nonlinear transformation is performed using the LeakyReLU activation function; The transformation result is normalized using the softmax function to obtain normalized attention coefficients, which are then used for weighted aggregation of neighbor node features.

8. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 1, characterized in that, The steps for training the network parameters include: initializing the network parameters, setting the optimizer, initial learning rate, maximum training steps and loss function, and iteratively training the network using the training graph data until the maximum training steps are reached or the loss function converges.

9. The planetary gearbox fault diagnosis method based on a physically enhanced graph attention network according to claim 1, characterized in that, The steps for applying constraints using a hybrid loss function include: Cross-entropy loss is used to measure the difference between the model's predicted category and the true label; The similarity between the distribution of attention coefficients in the KL divergence loss constrained graph attention network and the distribution of the integrated edge relation weights in the training graph data is calculated. The cross-entropy loss and the KL divergence loss are weighted and summed to form the final mixed loss function.

Images