A method and device for diagnosing gearbox faults using graph convolutional neural network

Through the combination of EMD decomposition and dual-channel graph convolutional neural network, a graph structure feature set is constructed, which solves the gearbox fault diagnosis problem with limited sample size, realizes efficient fault identification and diagnosis, and reduces the computing cost.

CN118776871BActive Publication Date: 2024-11-12KUNMING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411258581.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-11-12
Estimated Expiration
2044-09-09

AI Technical Summary

Technical Problem

It is difficult to effectively construct graph feature structure data sets in the prior art, especially when the sample size is limited, it is difficult to improve the accuracy of graph neural networks for diagnosis of gearbox failures.

Method used

IMF modal components are obtained through EMD decomposition, and graph structural feature sets are constructed by combining time domain and frequency domain features, and fault diagnosis is performed using a dual-channel graph convolution neural network and attention mechanism, reducing operation costs and improving diagnostic accuracy.

Benefits of technology

On the premise of ensuring diagnostic accuracy, using less data to obtain more graph structure feature data samples, reducing computing costs, and providing a new fault diagnosis technology route and strategy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118776871B_ABST
    Figure CN118776871B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and device for diagnosing gearbox faults using a graph convolutional neural network, and belongs to the field of fault diagnosis of mechanical equipment. The method includes: constructing a gearbox fault diagnosis data set; decomposing data samples in the gearbox fault diagnosis data set into IMF modal components with the optimal number of modal modes, and then performing FFT transformation to obtain the frequency domain feature representation of the IMF modal components; extracting the time domain features and frequency domain features of each IMF modal component to construct a feature node set; combining the time domain features and frequency domain features of each IMF modal component as nodes to construct a graph structure feature set; using the vibration data of the gearbox to be diagnosed as the input of a dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state. Compared with the prior art, the method of the present invention can obtain relatively more graph structure feature data samples using less data while ensuring the accuracy of diagnosis, and can greatly reduce the computing cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for diagnosing gear box faults by using a graph convolutional neural network, and belongs to the field of fault diagnosis of mechanical equipment. Background Art

[0002] Gearboxes are important parts of various electromechanical equipment, including rotating machinery such as gears and bearings. They have the functions of bearing and fixing shaft systems, transmitting speed, torque, and energy. Due to the various alternating loads and continuous power transmission, important parts such as gears and bearings in gearboxes are prone to aging and damage. Therefore, how to efficiently and accurately detect faults in gearboxes is of great significance in improving production efficiency, ensuring property safety, protecting personal safety, and maintaining normal production operations.

[0003] Graph neural network is a neural network model designed for processing non-Euclidean structure data. In practice, various non-Euclidean data problems can be abstracted into graph structures. Graph neural network provides a potential solution for problems such as gearbox fault detection that require processing non-Euclidean structure data. Using graph neural network, the connection and operating status between these components can be modeled and learned. By training these graph data, the model can learn the normal interaction patterns between the various components, so that abnormal conditions and potential fault signals can be detected.

[0004] One of the difficulties in using graph neural networks for fault diagnosis is to construct a suitable graph feature structure dataset. To ensure that the graph feature structure data is easy to construct and the network can fully learn the vector representations of nodes and edges at the same time, improving the accuracy of model diagnosis when the sample size is limited becomes a problem that needs to be solved.

[0005] In view of this, the present invention is proposed. Summary of the invention

[0006] Based on the problems in the above background technology, the present invention provides a method and device for diagnosing gearbox faults using a graph convolutional neural network, so as to diagnose gearbox faults.

[0007] The technical solution of the present invention is:

[0008] According to a first aspect of the present invention, a method for diagnosing gearbox faults using a graph convolutional neural network is provided, comprising: a data acquisition and preprocessing step: acquiring gearbox vibration data under different states, and preprocessing the gearbox vibration data under different states respectively for constructing a gearbox fault diagnosis data set; wherein the state includes a healthy state and a fault state; a feature node extraction step: based on a preset optimal number of modes, using EMD to decompose data samples in the gearbox fault diagnosis data set, decomposing the data samples in the gearbox fault diagnosis data set into IMF modal components with an optimal number of modes having different center frequencies; then performing an FFT transformation on the decomposed IMF modal components to obtain a frequency domain feature representation of the IMF modal components; extracting the time domain features and frequency domain features of each IMF modal component to construct a feature node set; and constructing a graph structure feature set. Construction steps: Combine the time domain features and frequency domain features of each IMF modal component as a node; For each node, construct the graph structure edge under the node based on the adjacent nodes; Determine the weight of each edge in the graph structure based on the Gaussian kernel weight to construct a graph structure feature set; Fault diagnosis model construction and training steps: Integrate the attention mechanism to construct a dual-channel graph convolutional neural network fault diagnosis model, use the graph structure feature set as the input of the dual-channel graph convolutional neural network fault diagnosis model, implement the training of the dual-channel graph convolutional neural network fault diagnosis model, and obtain the trained dual-channel graph convolutional neural network fault diagnosis model; Gearbox fault diagnosis steps: Use the vibration data of the gearbox to be diagnosed to construct a graph structure feature set as the input of the trained dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state and diagnose the gearbox fault.

[0009] In the process of constructing the graph structure feature set, the KNN algorithm is used to mine the node association relationship to find the The number of samples of nearest neighbors , get the node's adjacent nodes.

[0010] Said The determination of the optimal number of modes is as follows: The initial value of the optimal mode number is set; in the In the case of taking the initial value, by successively increasing the value of the optimal number of modes, and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the value of the optimal number of modes; after determining the value of the optimal number of modes, according to the determined optimal number of modes, in the Within the preset range of the optimal value of When taking different integer values, by increasing the The value of , and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the The best value of .

[0011] The dual-channel graph convolutional neural network fault diagnosis model specifically includes: an input layer, two graph attention network channels, a fully connected layer, a Dropout layer and an output layer; wherein the input layer serves as the input of the two graph attention network channels, the outputs of the two graph attention network channels are concatenated as the input of the fully connected layer, and are passed through a Dropout layer as the input of the output layer, and finally the output layer outputs the state category of the gearbox vibration data; the two graph attention network channels have the same structure, both including two stacked graph attention convolutional layers, and a BatchNorm layer and a Relu activation function are introduced after each graph attention convolutional layer.

[0012] According to a second aspect of the present invention, there is provided a device for diagnosing gearbox faults using a graph convolutional neural network, comprising: a data acquisition and preprocessing module: used to acquire gearbox vibration data under different states, and preprocess the gearbox vibration data under different states respectively for constructing a gearbox fault diagnosis data set; wherein the state includes a healthy state and a fault state; a feature node extraction module: used to decompose data samples in the gearbox fault diagnosis data set using EMD according to a preset optimal number of modes, and decompose the data samples in the gearbox fault diagnosis data set into IMF modal components with an optimal number of modes with different center frequencies; then perform FFT transformation on the decomposed IMF modal components to obtain frequency domain feature representations of the IMF modal components; extract time domain features and frequency domain features of each IMF modal component to construct a feature node set; graph structure feature set construction Module: used to combine the time domain features and frequency domain features of each IMF modal component as a node; for each node, the graph structure edge under the node is constructed based on the adjacent nodes; the weight of each edge in the graph structure is determined based on the Gaussian kernel weight, so as to construct a graph structure feature set; fault diagnosis model construction and training module: used to integrate the attention mechanism to construct a dual-channel graph convolutional neural network fault diagnosis model, use the graph structure feature set as the input of the dual-channel graph convolutional neural network fault diagnosis model, realize the training of the dual-channel graph convolutional neural network fault diagnosis model, and obtain the trained dual-channel graph convolutional neural network fault diagnosis model; gearbox fault diagnosis module: used to use the vibration data of the gearbox to be diagnosed to construct a graph structure feature set as the input of the trained dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state and realize the diagnosis of the gearbox fault.

[0013] The beneficial effects of the present invention are as follows: the present invention obtains training sets and test sets by constructing a graph structure feature set, and on this basis builds a dual-channel graph convolutional neural network fault diagnosis model constructed by integrating an attention mechanism to evaluate the state of the gearbox to achieve fault diagnosis; compared with the prior art, the method of the present invention can use less data to obtain relatively more graph structure feature data samples while ensuring the accuracy of diagnosis, and can greatly reduce the computing cost; at the same time, the method provides researchers in the field of fault diagnosis with a new technical route and implementation strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the process of the present invention;

[0015] Figure 2 A flowchart for constructing a graph feature structure data set of the present invention;

[0016] Figure 3 It is a schematic diagram of the structure of the dual-channel graph convolutional neural network fault diagnosis model of the present invention;

[0017] Figure 4 Visualization of model training results. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.

[0019] Example 1: Figure 1-4As shown, according to a first aspect of an embodiment of the present invention, a method for diagnosing gearbox faults using a graph convolutional neural network is provided, comprising: a data acquisition and preprocessing step: acquiring gearbox vibration data under different states, and preprocessing the gearbox vibration data under different states respectively for constructing a gearbox fault diagnosis data set; wherein the state includes a healthy state and a fault state; a feature node extraction step: based on a preset optimal number of modes, using EMD to decompose data samples in the gearbox fault diagnosis data set, decomposing the data samples in the gearbox fault diagnosis data set into IMF modal components with an optimal number of modes having different center frequencies; then performing FFT transformation on the decomposed IMF modal components to obtain frequency domain feature representations of the IMF modal components; extracting time domain features and frequency domain features of each IMF modal component to construct a feature node set; graph structure features Collection and construction steps: combine the time domain features and frequency domain features of each IMF modal component as a node; for each node, construct the graph structure edge under the node based on the adjacent nodes; determine the weight of each edge in the graph structure based on the Gaussian kernel weight, so as to construct a graph structure feature set; fault diagnosis model construction and training steps: integrate the attention mechanism to construct a dual-channel graph convolutional neural network fault diagnosis model, use the graph structure feature set as the input of the dual-channel graph convolutional neural network fault diagnosis model, realize the training of the dual-channel graph convolutional neural network fault diagnosis model, and obtain the trained dual-channel graph convolutional neural network fault diagnosis model; gearbox fault diagnosis steps: use the vibration data of the gearbox to be diagnosed to construct a graph structure feature set as the input of the trained dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state and realize the diagnosis of gearbox faults.

[0020] Furthermore, in the process of constructing the graph structure feature set, the KNN algorithm is used to mine the node association relationship to find the The number of samples of nearest neighbors , get the node's adjacent nodes.

[0021] Furthermore, the The determination of the optimal number of modes is as follows: The initial value of the optimal mode number is set; in the In the case of taking the initial value, by successively increasing the value of the optimal number of modes, and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the value of the optimal number of modes; after determining the value of the optimal number of modes, according to the determined optimal number of modes, in the Within the preset range of the optimal value of When taking different integer values, by increasing the The value of , and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the The best value of .

[0022] Furthermore, the dual-channel graph convolutional neural network fault diagnosis model specifically includes: an input layer, two graph attention network channels, a fully connected layer, a Dropout layer and an output layer; wherein the input layer serves as the input of the two graph attention network channels, the outputs of the two graph attention network channels are concatenated as the input of the fully connected layer, and after passing through a Dropout layer, they are used as the input of the output layer, and finally the output layer outputs the state category of the gearbox vibration data; the two graph attention network channels have the same structure, both including two stacked graph attention convolutional layers, and a BatchNorm layer and a Relu activation function are introduced after each graph attention convolutional layer.

[0023] The optimal number of modes, are all integers. In the embodiments of the present invention, The preset interval of the optimal value is [1,4]; The initial value of is 3; the initial value of the optimal number of modes is 3, and the value of the optimal number of modes is increased in sequence (the step size is 1, such as after the initial value is 3, then 4, and so on), and the graph structure feature set construction method is used to input the graph structure feature set into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy; it should be noted that during the training process, as the number of optimal modes increases, the training accuracy increases first and then decreases.

[0024] Determined by the above method The optimal value of the number of neighbor samples can, on the one hand, avoid the problem that it is too small, resulting in each node in the graph being connected to only a few nearest neighbor nodes, and it can avoid the problem that it is too small, resulting in the network focusing too much on local information, and the information propagation between nodes is restricted, which makes it difficult for the network to effectively summarize useful information from different parts; on the other hand, it can avoid being too large, resulting in each node in the graph being connected to many nodes, which may introduce a lot of information that is irrelevant to the current node or has weak correlation, thereby increasing noise and reducing the network's discrimination ability; at the same time, it can avoid being too large, resulting in an increase in the density of the graph, which makes the information transmission between nodes more extensive, resulting in the network's processing of all nodes tending to be averaged and weakening the sensitivity to specific failure modes. Furthermore, regarding the number of IMF modal components obtained by EMD decomposition, if the optimal number of modes is too small, EMD decomposition may not be able to fully extract all frequency components in the signal, especially high-frequency components related to minor faults, which will cause some key fault features to be ignored, thereby reducing the sensitivity and accuracy of fault detection; at the same time, when the optimal number of modes is too small, different frequency components may be mixed in the same IMF modal component, resulting in these IMF modal components being difficult to accurately reflect the specific patterns in the signal; if the optimal number of modes is too large, EMD may decompose the noise components in the original signal into independent IMF modal components. These IMF modal components containing noise will increase the difficulty of feature extraction. Too many IMF modal components will also cause multiple IMF modal components extracted from the same signal to contain repeated information. This redundant information will make it more difficult for the model to focus on truly useful fault features, thereby weakening the performance of the diagnostic model. In view of the fact that the length of the data segment taken in this paper is 1024 data points, the number of optimal modes is initially set to 3-7. In addition, the number of IMF modal components obtained by decomposing each segment of data is used as the number of nodes to construct the graph. Therefore, combined with the characteristics of the KNN algorithm, The initial value of the number of neighbor samples is set to 3 for the experiments.

[0025] One difficulty in using graph neural networks for fault diagnosis is to construct a suitable graph feature structure data set. In order to ensure that the graph feature structure data is easy to construct and the network can fully learn the vector representation of nodes and edges at the same time, the accuracy of model diagnosis can be improved under the condition of limited sample size. The present invention uses EMD decomposition to construct the graph structure feature set, but the construction using raw data will increase the computational cost. Therefore, the present invention introduces the use of EMD decomposition to construct the graph structure feature set based on the preset optimal number of modes, which can decompose the original signal into multiple IMF modal components, and extract the time domain and frequency domain features of the IMF modal components on this basis to achieve data dimension reduction and reduce computational costs. Based on this, the present invention further uses the time domain and frequency domain features extracted from each layer as nodes, and uses the KNN algorithm to find the nearest neighbor of each node and calculate the weight of the edge, thereby constructing a graph structure feature set; finally, the dual-channel graph convolutional neural network fault diagnosis model constructed by integrating the attention mechanism is used for state recognition, so as to perform fault diagnosis under limited data.

[0026] According to a second aspect of an embodiment of the present invention, there is provided a device for diagnosing gearbox faults using a graph convolutional neural network, comprising: a data acquisition and preprocessing module: used to acquire gearbox vibration data under different states, and preprocess the gearbox vibration data under different states respectively for constructing a gearbox fault diagnosis data set; wherein the state includes a healthy state and a fault state; a feature node extraction module: used to decompose data samples in the gearbox fault diagnosis data set using EMD according to a preset optimal number of modes, and decompose the data samples in the gearbox fault diagnosis data set into IMF modal components with an optimal number of modes with different center frequencies; then perform FFT transformation on the decomposed IMF modal components to obtain frequency domain feature representations of the IMF modal components; extract time domain features and frequency domain features of each IMF modal component to construct a feature node set; graph structure feature set construction Construction module: used to combine the time domain features and frequency domain features of each IMF modal component as a node; for each node, the graph structure edge under the node is constructed based on the adjacent nodes; the weight of each edge in the graph structure is determined based on the Gaussian kernel weight, so as to construct a graph structure feature set; fault diagnosis model construction and training module: used to integrate the attention mechanism to construct a dual-channel graph convolutional neural network fault diagnosis model, use the graph structure feature set as the input of the dual-channel graph convolutional neural network fault diagnosis model, realize the training of the dual-channel graph convolutional neural network fault diagnosis model, and obtain the trained dual-channel graph convolutional neural network fault diagnosis model; gearbox fault diagnosis module: used to use the vibration data of the gearbox to be diagnosed to construct a graph structure feature set as the input of the trained dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state and realize the diagnosis of the gearbox fault.

[0027] Embodiment 2: Combined with experimental data, the optional specific implementation process of the present invention is described as follows:

[0028] like Figure 1-4 As shown, a method for diagnosing gearbox faults using a graph convolutional neural network includes:

[0029] (1) Raw data acquisition and preprocessing

[0030] This embodiment takes the public PHM2024 competition data set as an example to introduce gearbox fault diagnosis. The test bench of the public PHM2024 competition data set includes a motor, a reduction gearbox, and an axle box. The gearbox fault types include tooth root cracks, tooth surface wear, missing teeth, broken teeth, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, and bearing cage fault. Three different speeds of 20HZ, 40HZ, and 60HZ, as well as three different loads of 0KN, 10KN, and -10KN are set. There are a total of 8 fault state signals and 1 healthy state signal for 9 classification tasks.

[0031] For the gearbox vibration data under 9 states involved in the public PHM2024 competition data set, one piece of gearbox vibration data under each of the 9 states is randomly selected as training data, and the segment length of the training data is set to 1024 data points. 300 original data samples can be obtained for each state; the original data samples under each state are normalized respectively, and the normalized data samples are used as the gearbox fault diagnosis data set.

[0032] (2) Feature node extraction

[0033] The segmented data is decomposed using EMD, and each segmented data is decomposed into IMF modal components with different center frequencies. Then, the first five IMF modal components obtained by each decomposition are transformed by FFT to obtain their frequency domain feature representation. Finally, the time domain features and frequency domain features of the first five IMF modal components are extracted, and the features are combined together to construct a feature node set.

[0034] When the number of IMF modal components is selected as 5, each data sample in the gearbox fault diagnosis data set can obtain 135 features (27 features of each IMF modal component as a node). Therefore, when the number of data samples in each state is 300, there are a total of 1500 feature node sets for each state.

[0035] In an embodiment of the present invention, the time domain features of the extracted IMF modal components include: mean, effective value, variance, standard deviation, maximum value, minimum value, peak value, peak-to-peak value, skewness, kurtosis, peak index, waveform index, pulse index and margin index, a total of 14 features; the frequency domain features of the extracted IMF modal components include: spectral mean, spectral variance, spectral skewness, spectral kurtosis, weighted average frequency, standard deviation of the spectrum, root mean square frequency of the spectrum, fourth-order center distance of the spectrum, ratio of the root mean square frequency of the frequency to the root mean square frequency of the spectrum amplitude, ratio of the standard deviation to the weighted average frequency, standardization of the skewness of the spectrum, standardization of the kurtosis of the spectrum, and standardization of the square root of the standard deviation of the spectrum, a total of 13 features.

[0036] (3) Construction of graph structure feature set

[0037] First, the time domain characteristics and frequency domain characteristics of each IMF modal component are combined as nodes; for each node, the structural edge of the node is constructed based on the adjacent nodes; the KNN algorithm is used to mine the node association relationship to find the The number of samples of the nearest neighbors, get the node The adjacent nodes are expressed as:

[0038] ;

[0039] in, Function returns a collection Midpoint Before nearest neighbors, in this embodiment of the present invention, Set to 3, Indicated by A node sample subset consisting of node samples, Representation Node The set of adjacent nodes.

[0040] Finally, the weight of each edge in the graph structure is determined based on the Gaussian kernel weight to construct the graph structure feature set. The Gaussian kernel weight is defined as follows:

[0041] ;

[0042] in, For Node and nodes The edge weights between is the bandwidth of the Gaussian kernel, and the graph structure feature set construction process is as follows Figure 2 shown.

[0043] In the embodiment of the present invention, the initial value of the optimal number of modes is first set to 3. The initial value of is 3. The preset interval of the optimal value of is [1,4]; In the case of taking the initial value, by successively increasing the optimal modal number of decomposition, and using the graph structure feature set construction step, the graph data is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the number of IMF modal components. The optimal modal number of the present invention is 5; after determining the value of the optimal modal number, according to the determined optimal modal number, Within the preset range of the optimal value of When taking different integer values, by increasing The value of , and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining The best value of is still 3.

[0044] In an embodiment of the present invention, the node dimension in each graph structure is 27, and 5 nodes construct a graph structure, that is, a normalized data sample can construct 1 graph structure feature data sample, so each state has a total of 300 graph structure feature data samples, and 9 states have a total of 2700 graph structure feature data samples.

[0045] (4) Fault diagnosis model construction and training

[0046] The dual-channel graph convolutional neural network fault diagnosis model constructed by integrating the attention mechanism is as follows Figure 3As shown in the figure, specifically: First, the input graph structure feature set passes through two channels respectively: the first channel first passes through the first graph convolution layer (GATConv) fused with the attention mechanism to capture the information transfer between direct neighbors, that is, the features of neighbor nodes are aggregated to the central node through the attention mechanism, and then passes through a BatchNorm layer to normalize the input of each mini-batch, so that the distribution of data in each layer of the neural network is more stable, reducing the problem of gradient disappearance or explosion, and accelerating the convergence of the model, and then passes through a ReLU activation function, thereby introducing nonlinearity, so that the graph network can capture the complex nonlinear relationship in the graph; second A graph convolution layer (GATConv) with an integrated attention mechanism is used to capture further relationships, which can help the model capture more complex relationships in the graph. Finally, a BatchNorm layer and a ReLU activation function are used to further accelerate the convergence of the model and capture complex nonlinear relationships in the graph. The second parallel channel is used to extract node and graph features from multiple perspectives to improve the model's ability to understand complex graph structures. By splicing the output features of the two channels, it is equivalent to fusing the features of two different perspectives, so that the subsequent decision layer can make judgments based on more comprehensive information, more easily adapt to task requirements, process different types of features, and enhance the versatility of the model. Finally, the fused features pass through a fully connected layer (FC), a regularization layer (Dropout), and an output layer. The fully connected layer (FC) is used to map the spliced ​​high-dimensional features to a new feature space, and the regularization layer (Dropout) is used to reduce the interdependence between neurons, thereby enhancing the generalization ability of the model and reducing the risk of overfitting. The output layer converts the features from the fully connected layer into the final prediction results. Further, such as Figure 3 As shown in the figure, both attention network channels include: GATConv1, BatchNorm layer, ReLU activation function, GATConv2, BatchNorm layer, ReLU activation function according to the secondary connection; GATConv1, output dimension is 128, GATConv2 output dimension is 128. The input dimension of the fully connected layer FC is 256, and the output dimension is 64; the dropout rate of the Dropout layer is set to 0.2; the input dimension of the output layer is 64, and the output dimension is 9.

[0047] The graph structure feature set is divided into 8:2, that is, the ratio of the training set is 0.8, the ratio of the test set is 0.2, the initial learning rate is set to 0.001, the parameters are updated through back propagation and Adam optimization algorithm, and the training process adopts mini-batch learning, and the size is set to 32. The loss function uses the cross entropy loss function. While training, the model is also tested with the test set until the training is completed, and the trained dual-channel graph convolutional neural network fault diagnosis model is obtained.

[0048] The network is trained with the gearbox vibration data in the embodiment of the present invention to obtain the best accuracy of the model, and the model output results are visualized and analyzed using the confusion matrix. The visualization results are as follows: Figure 4 As shown; 0 represents the healthy state of the gearbox, 1 represents the crack of the gearbox tooth root, 2 represents the wear of the gearbox tooth surface, 3 represents the missing teeth of the gearbox, 4 represents the broken teeth of the gearbox, 5 represents the inner ring fault of the bearing, 6 represents the outer ring fault of the bearing, 7 represents the rolling element fault of the bearing, and 8 represents the bearing cage fault; in the confusion matrix, the horizontal axis represents the predicted label of each state of the gearbox, and the vertical axis represents the true label of each state of the gearbox. The numerical value represents the proportion of all correctly predicted samples to the total number. The matrix elements are concentrated on the diagonal, which means that the diagnostic effect of the algorithm is better. Figure 4 It can be concluded that the model can basically accurately identify the label of each sample, and only a few samples of tooth surface wear and bearing outer ring failure are judged as missing teeth and healthy status.

[0049] At the same time, the gearbox vibration data is input into different models to obtain the best accuracy of the model and compare it with the method of the present invention to verify the superiority of the present invention. In the embodiment of the present invention, the models selected for comparison are: a single-channel graph convolutional neural network (i.e., GCN) under a graph structure feature set directly constructed by the original data, a single-channel graph convolutional neural network (i.e., GAT) with a fusion attention mechanism under a graph structure feature set directly constructed by the original data, a single-channel graph convolutional neural network (i.e., EMD+GCN) under a graph structure feature set constructed by extracting features from the original data after EMD decomposition, a single-channel graph attention network (EMD+GAT) with a fusion attention mechanism under a graph structure feature set constructed by extracting features from the original data after EMD decomposition, and the present invention (i.e., EMD+BiGAT).

[0050] The dual-channel graph convolutional neural network fault diagnosis model constructed by integrating the attention mechanism and the comparison model were trained 100 times according to the above parameter settings to obtain the experimental results shown in Table 1.

[0051] Table 1 Model experimental results

[0052]

[0053] It can be seen from the experimental results of Table 1 in the embodiment of the present invention that by comparing the experimental results of GCN and GAT, adding an attention mechanism to the model can effectively capture useful feature information in the graph structure, thereby improving the accuracy of the model; by comparing the model accuracy of GCN and EMD+GCN, it can be seen that by using EMD decomposition, the number of graph structure feature sets can be increased under the condition of limited original data, and the accuracy of the model can be effectively improved when the node feature dimension is small; by comparing the experimental results of EMD+GAT and EMD+BiGAT, it can be seen that the dual channel can extract the feature information of nodes and graphs from different perspectives, thereby improving the model's ability to understand complex graph structures. This shows that the EMD+BiGAT of the present invention can effectively improve the accuracy of the method of the present invention by increasing the number of graph structure feature data samples and using the technical solution of the dual-channel graph convolutional neural network that integrates the attention mechanism, fully proving the superiority of the present invention.

[0054] (5) Gearbox fault identification and diagnosis

[0055] Based on the trained dual-channel graph convolutional neural network fault diagnosis model, the gearbox vibration data that did not participate in the training was preprocessed and feature nodes were extracted in sequence, and then the graph structure feature set was constructed. Finally, it was input into the trained dual-channel graph convolutional neural network fault diagnosis model for gearbox state recognition and gearbox fault diagnosis.

[0056] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for diagnosing gearbox faults using graph convolutional neural networks. It is characterized in that include: Data acquisition and preprocessing steps: acquiring gearbox vibration data under different states, and preprocessing the gearbox vibration data under different states respectively to construct a gearbox fault diagnosis data set; wherein the states include healthy state and fault state; Feature node extraction steps: according to the preset optimal number of modes, use EMD to decompose the data samples in the gearbox fault diagnosis data set, and decompose the data samples in the gearbox fault diagnosis data set into IMF modal components with different optimal modal numbers of center frequencies; then perform FFT transformation on the decomposed IMF modal components to obtain the frequency domain feature representation of the IMF modal components; extract the time domain features and frequency domain features of each IMF modal component to construct a feature node set; The steps of constructing the graph structure feature set are as follows: the time domain features and frequency domain features of each IMF modal component are combined as nodes; for each node, the graph structure edges under the node are constructed based on the adjacent nodes; the weight of each edge in the graph structure is determined based on the Gaussian kernel weight, so as to construct the graph structure feature set; Fault diagnosis model construction and training steps: Integrate the attention mechanism to construct a dual-channel graph convolutional neural network fault diagnosis model, use the graph structure feature set as the input of the dual-channel graph convolutional neural network fault diagnosis model, implement the training of the dual-channel graph convolutional neural network fault diagnosis model, and obtain the trained dual-channel graph convolutional neural network fault diagnosis model; Gearbox fault diagnosis steps: Use the vibration data of the gearbox to be diagnosed to construct a graph structure feature set as the input of the trained dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state and diagnose the gearbox fault; In the process of constructing the graph structure feature set, the KNN algorithm is used to mine the node association relationship to find the The number of samples of nearest neighbors , get the node's adjacent nodes; Said The determination of the optimal number of modes is as follows: The initial value of the optimal mode number is set; in the In the case of taking the initial value, by successively increasing the value of the optimal number of modes, and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the value of the optimal number of modes; after determining the value of the optimal number of modes, according to the determined optimal number of modes, in the Within the preset range of the optimal value of When taking different integer values, by increasing the The value of , and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the The best value of ; The dual-channel graph convolutional neural network fault diagnosis model specifically includes: an input layer, two graph attention network channels, a fully connected layer, a Dropout layer and an output layer; wherein the input layer serves as the input of the two graph attention network channels, the outputs of the two graph attention network channels are concatenated as the input of the fully connected layer, and are passed through a Dropout layer as the input of the output layer, and finally the output layer outputs the state category of the gearbox vibration data; the two graph attention network channels have the same structure, both including two stacked graph attention convolutional layers, and a BatchNorm layer and a Relu activation function are introduced after each graph attention convolutional layer.

2. A device for diagnosing gearbox faults using graph convolutional neural networks, It is characterized in that include: Data acquisition and preprocessing module: used to acquire gearbox vibration data under different states, and preprocess the gearbox vibration data under different states respectively to construct a gearbox fault diagnosis data set; wherein the states include healthy state and fault state; Feature node extraction module: used to decompose the data samples in the gearbox fault diagnosis data set using EMD according to the preset optimal number of modes, and decompose the data samples in the gearbox fault diagnosis data set into IMF modal components with different optimal numbers of central frequencies; then perform FFT transformation on the decomposed IMF modal components to obtain the frequency domain feature representation of the IMF modal components; extract the time domain features and frequency domain features of each IMF modal component to construct a feature node set; Graph structure feature set construction module: used to combine the time domain features and frequency domain features of each IMF modal component as nodes; for each node, the graph structure edge under the node is constructed based on the adjacent nodes; the weight of each edge in the graph structure is determined based on the Gaussian kernel weight, so as to construct the graph structure feature set; Fault diagnosis model construction and training module: used to integrate the attention mechanism to build a dual-channel graph convolutional neural network fault diagnosis model, use the graph structure feature set as the input of the dual-channel graph convolutional neural network fault diagnosis model, implement the training of the dual-channel graph convolutional neural network fault diagnosis model, and obtain the trained dual-channel graph convolutional neural network fault diagnosis model; Gearbox fault diagnosis module: used to construct a graph structure feature set using the vibration data of the gearbox to be diagnosed as the input of the trained dual-channel graph convolutional neural network fault diagnosis model to identify the gearbox state and diagnose the gearbox fault; In the process of constructing the graph structure feature set, the KNN algorithm is used to mine the node association relationship to find the The number of samples of nearest neighbors , get the node's adjacent nodes; Said The determination of the optimal number of modes is as follows: The initial value of the optimal mode number is set; in the In the case of taking the initial value, by successively increasing the value of the optimal number of modes, and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the value of the optimal number of modes; after determining the value of the optimal number of modes, according to the determined optimal number of modes, in the Within the preset range of the optimal value of When taking different integer values, by increasing the The value of , and using the graph structure feature set construction step, the graph structure feature set is input into the dual-channel graph convolutional neural network fault diagnosis model to obtain the best training accuracy, thereby determining the The best value of ; The dual-channel graph convolutional neural network fault diagnosis model specifically includes: an input layer, two graph attention network channels, a fully connected layer, a Dropout layer and an output layer; wherein the input layer serves as the input of the two graph attention network channels, the outputs of the two graph attention network channels are concatenated as the input of the fully connected layer, and are passed through a Dropout layer as the input of the output layer, and finally the output layer outputs the state category of the gearbox vibration data; the two graph attention network channels have the same structure, both including two stacked graph attention convolutional layers, and a BatchNorm layer and a Relu activation function are introduced after each graph attention convolutional layer.