Rotating machine fault diagnosis method based on multi-view space-time diagram attention network

Abstract

The invention discloses a rotating machine fault diagnosis method based on a multi-view space-time diagram attention network, and belongs to the technical field of rotating machine intelligent monitoring and fault diagnosis. A multi-view space-time diagram attention network sequentially comprises a topological graph construction layer, a two-channel parallel diagram attention network, a multi-view feature fusion module, a gating recursive unit time feature extraction module, a full connection layer and a Softmax function layer, and during training, firstly, a distance topological graph and a similarity topological graph of input data are constructed; through a two-channel parallel graph attention network, spatial feature extraction is carried out from two perspectives of geometric distance and feature similarity; through a multi-view feature fusion module, dynamic fusion of dual-channel features is realized through a learnable attention weight matrix; and finally, the fused spatial features are input into the gated loop unit network, fault feature information in the time dimension is deeply mined, and the accuracy and robustness of fault diagnosis of the rotating machine can be effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent monitoring and fault diagnosis of rotating machinery, and particularly relates to a rotating machinery fault diagnosis method based on a multi-view spatiotemporal graph attention network. BACKGROUND

[0002] With the continuous rise of the intelligent degree of industrial equipment, the state monitoring of key components of rotating machinery is facing new technical challenges and requirements. As the core components of the transmission system, gears and bearings bear complex loads and friction during long-term operation and are prone to damage. Typical fault modes include but are not limited to gear tooth peeling, gear root cracking, rolling element pitting, inner and outer ring wear, etc. These mechanical damage problems not only significantly reduce the operating efficiency of mechanical equipment, but also greatly reduce the stability of equipment operation, and in severe cases, can even cause unexpected production interruptions, thereby causing huge economic losses to enterprises.

[0003] In actual industrial environments, traditional rotating machinery fault diagnosis methods are based on the analysis of vibration signals collected by sensors. Due to signal attenuation caused by transmission path effects, signal-to-noise ratio degradation caused by environmental noise, and feature aliasing caused by multi-source vibration coupling, traditional diagnosis methods have obvious shortcomings: on the one hand, signal analysis methods based on fast Fourier transform have limited sensitivity to minor damage; on the other hand, diagnosis systems relying on expert experience cannot cope with feature changes under different operating conditions. In addition, due to the explosive growth of data volume, traditional signal analysis-based diagnosis methods have been unable to meet the application requirements of rapid diagnosis.

[0004] Under the background of big data analysis, intelligent diagnosis methods represented by convolutional neural networks (CNN) have been widely applied. By constructing an end-to-end fault diagnosis model, not only can the dependence on professional knowledge be reduced, but also deep features in monitoring data can be fully mined. However, existing related network models mostly analyze and process data itself, and do not consider the interaction between multiple data, so the fault detection accuracy needs to be improved. SUMMARY

[0005] The present application is to solve the above technical problems existing in the prior art, and proposes a mechanical fault diagnosis method based on a multi-view spatiotemporal graph attention network.

[0006] The technical solution of the present application is a mechanical fault diagnosis method based on a multi-view spatiotemporal graph attention network model. The collected mechanical vibration signals to be tested are input into the network model for fault diagnosis. The network model is a multi-view spatiotemporal graph attention network model, which is constructed according to the following steps: Step 1. Collect vibration signals of rotation, including normal conditions and typical mechanical fault modes; Step 2. Preprocess the vibration signal and divide it into training and validation datasets; Step 3. Establish a multi-view spatiotemporal graph attention network and set network training conditions. The multi-view spatiotemporal graph attention network, from input to output, consists of a topology layer, a dual-channel parallel graph attention network, a multi-view feature fusion module, a gated recursive unit time feature extraction module, a fully connected layer, and a Softmax function layer. The topology layer is constructed by treating each fault data sample as a node in the topology graph, and constructing a distance topology graph and a similarity topology graph respectively. The dual-channel parallel graph attention network has the same architecture but does not share parameters. The constructed distance topology graph and similarity topology graph are respectively input to one channel of the dual-channel parallel graph attention network to extract distance view features and similarity view features respectively. The multi-view feature fusion module dynamically fuses the distance view features and similarity view features to obtain fused features. The gated recursive unit time feature extraction module further extracts time dimension features from the obtained fused features to obtain time dimension features. Step 4. Use the training dataset and validation dataset to train the multi-view spatiotemporal graph attention network to obtain the multi-view spatiotemporal graph attention network model.

[0007] The preferred step 2 involves using a sliding window method to segment the vibration signal into equal-length segments and perform data normalization to form data samples of each category. Then, according to the principle of stratified random sampling, the training set and validation set are randomly divided.

[0008] The preferred step 4 is as follows: Step 4.1 By constructing a topology layer, each faulty data sample is regarded as a node in the topology graph. Distance topology graph and similarity topology graph are constructed respectively. The specific operations are as follows: Step 4.1.1 Obtain the Euclidean distance matrix between all nodes from this perspective by measuring the Euclidean distance between any two nodes in the physical space. ; ; ; in, N Indicates the number of nodes used to construct the graph structure. Any two nodes , The Euclidean distance between them; Step 4.1.2 For the Euclidean distance matrix Perform KNN sparsification to retain each node.x i the front k shortest distance connections, forming a distance adjacency matrix and establishing a distance topology graph X , ); ; ; ; wherein, the distance metric neighborhood of a node , represents the edge connection weight between any two nodes based on distance metric, is the kernel function width; Step 4.1.3 obtains the similarity matrix of all nodes in this perspective by measuring the cosine similarity between node features ; ; ; wherein, represents the cosine value of the vector inner product angle between two nodes , ; Step 4.1.4 finds the nearest neighbors of each node k in the perspective of similarity using the KNN method, obtains the similarity adjacency matrix and constructs a similarity topology graph X , ); ; ; ; wherein, the similarity metric neighborhood of a node , represents the similarity adjacency matrix; Step 4.2 extracts the deep representation of nodes in different metric spaces by a double-channel parallel graph attention network, and the specific operation is as follows: Step 4.2.1 initializes the double-channel parallel graph attention network, and inputs the distance topology graph (X, ) and the similarity topology graph (X, ) into one channel of the double-channel parallel graph attention network respectively; Step 4.2.2: calculate and using Softmax the attention score of each edge in the adjacency matrix is obtained by normalization processing of the function; ; wherein || represents a node feature splicing operation, W is a shared weight matrix, represents a scoring function, LeakyReLU represents an activation function, represents the neighborhood nodes connected to the node in the graph; Step 4.2.3. Based on the attention score of each edge in the adjacency matrix, the aggregation of the node features is performed, and the distance perspective feature and the similarity perspective feature are obtained through the graph attention convolution operation, respectively. Step 4.3. Through the multi-perspective feature fusion module, the distance perspective feature and the similarity perspective feature are spliced along the feature dimension to generate the joint feature : ; The joint feature is self-attention fused to learn the interaction importance of different feature dimensions to obtain the fused feature . ; wherein, is a projection matrix, d is the dimension of the attention space; Step 4.4. The fused feature is input into the gated recurrent unit time feature extraction module to further extract the feature in the time dimension to obtain the time dimension feature . Step 4.5. The time dimension feature is input into the fully connected layer and processed through the Softmax function layer to obtain the output result output . ; wherein, respectively represent the weight and bias of the fully connected layer; Step 4.6. The cross-entropy loss of the output result output is calculated using the cross-entropy loss function, and it is determined whether the network is continuously iterated for several times without the cross-entropy loss decreasing or reaching the maximum iteration number. If not, return to step 4.1 to continue training; if yes, output the trained network model, i.e. the multi-perspective spatio-temporal graph attention network model.

[0009] The application constructs a multi-view spatio-temporal graph attention network from input to output in turn as a topological graph layer, a double-channel parallel graph attention network (GAT), a multi-view feature fusion module, a gated recurrent unit time feature extraction module (GRU), a full connection layer and a Softmax function layer, wherein, during network training, firstly, distance topological graphs and similarity topological graphs of input data are constructed from different measurement perspectives, spatial feature extraction is carried out from two perspectives of geometric distance and feature similarity through the double-channel parallel graph attention network; then, dynamic fusion of double-channel features is realized through a learnable attention weight matrix through the multi-view feature fusion module; finally, the fused spatial features are input into the gated recurrent unit network to deeply mine fault feature information in the time dimension.

[0010] Compared with the prior art, the application has the following beneficial effects: (1) The strategy of generating distance topological graphs and similarity topological graphs respectively breaks through the limitation of traditional single-perspective analysis, enables the network model to comprehensively capture fault feature information from different measurement perspectives such as geometric distance and feature similarity, and improves the representation integrity and accuracy of fault information, thereby laying a solid foundation for accurate fault diagnosis; (2) The combination of the double-channel parallel graph attention network (GAT) and the self-attention fusion module can precisely adjust the weight contribution of each perspective feature through an adaptive learning process according to the inherent characteristics and mutual correlation of different perspective features, thereby improving the adaptability of the fault diagnosis model to different data distributions and fault modes; (3) By cascading the gated recurrent unit time feature extraction module (GRU) and the double-channel graph attention network (GAT), the spatial correlation features under different perspectives are deeply mined from the topological structure by means of the powerful spatial feature extraction capability of the graph GAT, the time sequence dependence relationship in the fault development process is effectively captured by taking advantage of the GRU in time series data processing, the fault information is fully and deeply represented from the spatial and time dimensions, the extraction and analysis capability of the model for fault features is improved, and the accuracy and robustness of the rotating machinery fault diagnosis are improved. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 The flowchart of the network model constructed for the embodiment of the application.

[0012] Figure 2 The structural schematic diagram of the network model for the embodiment of the application.

[0013] Figure 3 The vibration data waveform diagram under different operating conditions of the embodiment of the application.

[0014] Figure 4A confusion matrix of a rotating machine fault diagnosis result of an embodiment of the application. DETAILED DESCRIPTION

[0015] A mechanical fault diagnosis method based on a multi-view spatio-temporal graph attention network model of the application is to input the collected mechanical vibration signal to be tested into the network model for fault diagnosis. The network model is a multi-view spatio-temporal graph attention network model, as shown in the following steps: Figure 1 Step 1. Collecting vibration signals of rotating machines containing normal state and typical fault modes; The rotating machine vibration test bench used in the embodiment of the application is composed of a driving motor, a transmission belt, a transmission shaft system, a load applicator, and a rolling bearing and other basic components. An acceleration sensor is arranged directly above the rolling bearing seat, and the sampling frequency is set to 12800 Hz. When the equipment runs to the rated speed / load working condition, vibration signals containing normal state and typical fault modes are collected. The specific typical fault mode is to manufacture crack defects on the outer ring, rolling body and inner ring of the cylindrical roller bearing of model N205EM. The bearing has a total of 13 rolling bodies, and the rolling body diameter is 7.5 mm , the pitch diameter of the bearing is 38.5 mm , and the contact angle is 0°. Step 2. Preprocessing the vibration signal and dividing the training data set and the verification data set; The vibration signal is segmented by the sliding window method and normalized, forming various types of data samples as shown in the following figure: Figure 3 After that, the training set and the verification set are randomly divided according to the stratified random sampling principle to ensure balanced distribution of each type of fault data sample. Figure 3 The data samples from top to bottom in the figure are normal state (no fault), bearing outer ring fault, bearing rolling body fault and bearing inner ring fault. The data is divided in a ratio of 8:2 in the embodiment of the application, that is, 80% of the data is randomly selected as the training data set, and the remaining 20% of the data is used as the verification data set. Step 3. Establishing a multi-view spatio-temporal graph attention network and setting network training conditions (setting model training times, optimizers, initial learning rate and other hyperparameters), as shown in the following figure: Figure 2 ​The shown from input to output is provided with a topology graph layer, a double-channel parallel graph attention network (GAT), a multi-view feature fusion module, a single-layer gated recurrent unit time feature extraction module (GRU), a 2-layer fully connected layer and a Softmax function layer in turn, the topology graph layer is to regard each fault data sample as a node in the topology graph, and distance topology graph and similarity topology graph are constructed respectively; the double channels of the double-channel parallel graph attention network have the same architecture but non-parameter sharing, the constructed distance topology graph and similarity topology graph are input into one channel of the double-channel parallel graph attention network respectively, and distance perspective feature and similarity perspective feature are obtained respectively; the multi-view feature fusion module fuses the distance perspective feature and the similarity perspective feature to obtain fusion features; the gated recurrent unit time feature extraction module further extracts features in the time dimension to obtain time dimension features; Step 4. Using the training data set and the validation data set, the multi-view spatio-temporal graph attention network is trained to obtain a multi-view spatio-temporal graph attention network model; Step 4.1 For each fault type, 10 fault data samples are randomly selected, each fault data sample is regarded as a node in the topology graph by constructing a topology graph layer, and distance topology graph and similarity topology graph are constructed respectively, the specific operation is as follows: Step 4.1.1 By measuring the Euclidean distance between any two nodes in the physical space, the Euclidean distance matrix between all nodes in this perspective is obtained ; ; ; Wherein, N represents the number of nodes for constructing the graph structure, is the Euclidean distance between any two nodes , ; Step 4.1.2 KNN sparse processing is performed on the Euclidean distance matrix , and the first shortest distance connection of each node k is retained to form a distance adjacency matrix and establish a distance topology graph X , ; ; ; ; Wherein, represents the distance measurement neighborhood of node , represents the edge connection weight between any two nodes based on distance metric, is the kernel function width; Step 4.1.3 obtains the similarity matrix of all nodes in this perspective by measuring the cosine similarity between node features ; ; ; wherein, represents the cosine value of the inner product angle between two nodes , ; Step 4.1.4 finds the K nearest neighbors of each node in the perspective of similarity by using the KNN method, obtains the similarity adjacency matrix k and constructs the similarity topology graph , X , ); ; ; ; wherein, represents the similarity metric neighborhood of node , represents the similarity adjacency matrix; Step 4.2 extracts the deep representation of nodes in different metric spaces by using the double-channel parallel graph attention network, and the specific operation is as follows: Step 4.2.1 initializes the double-channel parallel graph attention network, and inputs the distance topology graph (X, ) and the similarity topology graph (X, ) into one channel of the double-channel parallel graph attention network, respectively; Step 4.2.2: calculates and uses the Softmax function to perform normalization processing to obtain the attention score of each edge in the adjacency matrix; ; wherein, || represents the node feature concatenation operation, W is a shared weight matrix, represents a scoring function, LeakyReLU represents an activation function, represents the neighborhood nodes connected with node in the graph; Step 4.2.3 Aggregates node features based on the attention score of each edge in the adjacency matrix, and obtains distance-viewpoint features through graph attention convolution operation. and similarity perspective features ; Step 4.3: Using the multi-view feature fusion module, perform distance view features along the feature dimension. and similarity perspective features Perform a concatenation operation to generate joint features. : ; joint features Self-attention fusion is performed to learn the interaction importance of different feature dimensions, resulting in fused features. ; ; in, Let be the projection matrix. d For the dimension of the attention space; Step 4.4 will fuse the features The input is fed into a single-layer gated recursive unit time feature extraction module for further time-dimensional feature extraction, yielding the time-dimensional features. ; Step 4.5 Incorporate time dimension features Input to the fully connected layer and through Softmax After the function layer processes the integers, it obtains the output result. output ; ; in, These represent the weights and biases of the fully connected layer, respectively. Step 4.6 uses the cross-entropy loss function to calculate the output result. output The cross-entropy loss is used to determine whether the network has reached the maximum number of iterations or no longer decreases after 10 consecutive iterations. If not, return to step 4.1 to continue training; if yes, output the trained network model, i.e., the multi-view spatiotemporal graph attention network model.

[0016] The confusion matrix of the final diagnostic result in this embodiment of the invention is as follows: Figure 4 As shown. Figure 4 In the diagram, 1 is in normal condition (no fault), 2 is a faulty outer ring, 3 is a faulty rolling element, and 4 is a faulty inner ring. From... Figure 4 As can be seen, in the diagnostic results of this embodiment of the invention, except for one rolling element fault which was incorrectly identified as an inner ring fault, all other data were correctly classified, and the overall accuracy of the diagnostic model reached 99.75%, verifying the effectiveness of the method proposed in this invention for diagnosing rotating machinery faults.

Claims

1. A mechanical fault diagnosis method based on a multi-view spatiotemporal graph attention network model, wherein the collected mechanical vibration signal to be tested is input into the network model for fault diagnosis, characterized in that: The network model is a multi-view spatio-temporal graph attention network model, which is constructed according to the following steps: Step 1. Collecting vibration signals of rotating machinery including normal state and typical fault modes; Step 2. Preprocessing the vibration signals, dividing the training data set and the verification data set; Step 3. Establishing a multi-view spatio-temporal graph attention network and setting network training conditions, the multi-view spatio-temporal graph attention network is sequentially provided with a topology graph layer, a double-channel parallel graph attention network, a multi-view feature fusion module, a gated recurrent unit time feature extraction module, a full connection layer and a Softmax function layer from input to output, the topology graph layer is to regard each fault data sample as a node in the topology graph, and distance topology graph and similarity topology graph are constructed respectively; the double-channel parallel graph attention network has the same architecture but non-parameter sharing, the constructed distance topology graph and similarity topology graph are input into one channel of the double-channel parallel graph attention network respectively, and distance view features and similarity view features are obtained respectively; the multi-view feature fusion module dynamically fuses the distance view features and the similarity view features to obtain fusion features; the gated recurrent unit time feature extraction module further extracts features in the time dimension from the obtained fusion features to obtain time dimension features; Step 4. Training the multi-view spatio-temporal graph attention network using the training data set and the verification data set to obtain the multi-view spatio-temporal graph attention network model.

2. The mechanical fault diagnosis method of the multi-view spatio-temporal graph attention network model according to claim 1, characterized in that The step 2 is to segment the vibration signals into equal length by using the sliding window method and to perform data normalization processing to form various types of data samples, and then to randomly divide the training set and the verification set according to the stratified random sampling principle.

3. The mechanical fault diagnosis method of the multi-view spatio-temporal graph attention network model according to claim 2, characterized in that The step 4 is specifically as follows: Step 4.1, by constructing a topology graph layer, regarding each fault data sample as a node in the topology graph, respectively constructing a distance topology graph and a similarity topology graph, the specific operation is as follows: Step 4.1.1 Obtain the Euclidean distance matrix between all nodes in this view by measuring the Euclidean distance between any two nodes in the physical space ; ； ； wherein, N represents the number of nodes used to construct the graph structure, is the Euclidean distance between any two nodes , . Step 4.1.2 Euclidean distance matrix Perform KNN sparsification, keep the top K shortest distance connections of each node k to form a distance adjacency matrix and build a distance topology graph X , ​ ； ； ； wherein, representative node distance metric neighborhood, denotes the edge connection weight between any two nodes based on the distance metric, is the kernel width; Step 4.1.3 Obtain the similarity matrix of all nodes in this view by measuring the cosine similarity between node features ; ； ； wherein represents the cosine of the angle between the inner product of the vectors of two nodes , . Step 4.1.4 For each node , find its k nearest neighbors by KNN method in the view of similarity , get the similarity adjacency matrix X , , and construct the similarity topology graph ； ； ； wherein, representative nodes of the similarity measure neighborhood, denotes the similarity adjacency matrix; Step 4.2, by the double-channel parallel graph attention network, extracting deep representations of nodes in different metric spaces, the specific operation is as follows: Step 4.2.1 initializes a dual-channel parallel graph attention network, inputs the distance topology graph (X, ) and the similarity topology graph (X, ) into one channel of the dual-channel parallel graph attention network respectively; Step 4.2.2: Calculate according to the following formula Use Softmax The function is normalized to obtain the attention score for each edge in the adjacency matrix; ； wherein || denotes a node feature concatenation operation, W is a shared weight matrix, denotes a scoring function, LeakyReLU denotes an activation function, denotes a neighborhood node connected to the node in the graph. Step 4.2.

3. The aggregation of node features based on the attention score of each edge in the adjacency matrix, through the graph attention convolution operation, respectively obtains the distance perspective feature and the similarity perspective feature ; Step 4.3, the multi-view feature fusion module performs a stitching operation on the distance view feature and the similarity view feature along the feature dimension to generate a joint feature . . ​ ； Joint features Self-attention fusion is performed to learn the interaction importance of different feature dimensions to obtain fused features ; ； wherein, is a projection matrix, d is the dimension of the attention space; Step 4.4 fusing features input to the gated recurrent unit time feature extraction module for further time dimension feature extraction to obtain time dimension features ; Step 4.5: Time dimension feature Input to the fully connected layer and pass through Softmax After the function layer processing, the output result is obtained output ; ； wherein, respectively denote the weights and biases of the fully connected layer. Step 4.6 calculates the output result by using a cross-entropy loss function output If the cross-entropy loss of the network is not decreased for several times or reaches the maximum iteration number, the method returns to step 4.1 to continue training; otherwise, the trained network model, i.e., the multi-view spatio-temporal graph attention network model, is output.

Images