Rolling bearing fault diagnosis method based on multi-scale graph transformer

Abstract

The present invention discloses a rolling bearing fault diagnosis method based on a multi-scale graph transformer, and proposes a new graph node feature aggregation model, which aggregates the feature information of each neighborhood of the graph node into the feature representation of the central node, and uses the transformer to improve the utilization rate of important node information to enhance the feature representation. The implementation process is as follows: (1) performing wavelet transform on the collected original bearing fault data to obtain the fault time-frequency representation node, and using the first k fault time-frequency representation nodes of each fault time-frequency representation node to construct an adjacency matrix; (2) aggregating the multi-scale feature information of the fault time-frequency representation node through a multi-scale feature aggregation module, and then constructing the fault time-frequency representation node structure encoding to save the structural information of the fault time-frequency representation node, and then inputting it into the transformer module for feature extraction; (3) using the SoftMax classifier to classify the extracted fault features. Experimental results show that the method of the present invention is an effective fault diagnosis method.

Description

Technical Field

[0001] The invention relates to a rolling bearing fault diagnosis method based on a multi-scale graph Transformer, and belongs to the field of fault diagnosis. Background Art

[0002] Rolling bearings are indispensable parts of rotating machinery and are known as the "joints of industry". At the same time, they are also the most vulnerable parts. As the running time increases, rolling bearings are more prone to failure. If these failures are not discovered and maintained in time, the safety and service life of mechanical equipment are likely to be reduced. Traditional bearing fault diagnosis methods perform well when processing small-scale data sets, but perform poorly when facing large-scale data sets. In addition, their limited feature extraction capabilities make it difficult to fully explore the potential relationships between fault data.

[0003] With the rapid development of computing power, a large number of machine learning methods for intelligent fault diagnosis have been developed. Among them, methods based on deep neural networks, such as deep belief networks, autoencoders, convolutional neural networks, and long short-term memory networks, have attracted widespread attention from academia and industry. Deep neural networks have achieved great success in mechanical fault diagnosis, but there are still some problems. We believe that deep neural networks can learn the correlation between input features. However, this self-learning process for feature representation fails to explicitly mine the relationship between signals. In mechanical fault diagnosis, the relationship of monitoring fault signals varies greatly with the change of machine health status, so modeling and learning the signal relationship is effective for machine fault diagnosis. For graph data, each node is associated with other nodes through some complex edges, and the edges are used to represent the relationship between data. Due to the special form of graph data, traditional deep neural networks cannot achieve the same performance on graph data as on Euclidean data.

[0004] As a branch of deep learning, graph neural networks have achieved excellent performance in non-Euclidean data, such as recommendation systems, link prediction, node classification, and protein structure inference. They can fully explore the relationship between nodes for feature extraction, which brings new opportunities for the development of rolling bearing fault diagnosis. Although graph neural networks can achieve good performance on non-Euclidean data, for traditional graph neural networks, it only aggregates the information of local neighbor nodes, similar to the fixed kernel in convolutional neural networks, which means that it can only aggregate the information of a fixed number of nodes around the faulty node and cannot fully extract fault information. Scholars use multi-receptive field graph convolutional networks for machine fault diagnosis, avoiding the singleness of fixed scales and the weakness of feature extraction. However, the following problems still exist: First, the feature extraction method of multi-receptive fields does not focus on more useful feature information, which will lead to mixed feature information; second, the current method cannot effectively handle fault diagnosis in complex environments such as noise and variable working conditions. Third, the current effectiveness of bearing fault diagnosis in complex environments such as noise and variable working conditions is insufficient. Summary of the invention

[0005] This paper proposes a rolling bearing fault diagnosis method based on a multi-scale graph transformer to solve the problem that the current fault diagnosis algorithm cannot effectively handle fault diagnosis in complex environments such as noise and variable working conditions. The specific method steps are as follows:

[0006] 1. Collect the original bearing fault data through the sensor, enhance the data through the sliding window, and obtain the enhanced fault data; obtain the two-dimensional fault time-frequency representation node v∈R containing both time domain and frequency domain information through continuous wavelet transform h×m , where h×m is the dimension of the node; expand the fault time-frequency representation nodes horizontally and fuse all nodes vertically to obtain the feature matrix S∈R of the fault time-frequency representation nodes n×d , where n is the number of nodes and d = h × m.

[0007] 2. By calculating the distance between two fault time-frequency representation nodes, the first k fault time-frequency representation nodes of each fault time-frequency representation node are used to construct an adjacency matrix to obtain the adjacency matrix A of the fault time-frequency representation node.

[0008] 3. The multi-scale graph Transformer fault diagnosis model is used to extract features and learn representations of fault time-frequency representation nodes. The specific implementation method is as follows:

[0009] (3a) Construct a multi-scale feature aggregation module to aggregate the multi-scale feature information of the fault time-frequency representation node. The specific implementation of this module is as follows:

[0010] For the fault time-frequency representation nodes v and u, let N k(v) = {u∈v|d(v,u)≤k} is the k-th order neighborhood of v, where d(v,u) represents the shortest path distance between v and u; define N 0 (v) = {v}, that is, when k is 0, the 0th-order neighborhood is the node’s own feature; the aggregation operator is used in the multi-scale feature aggregation module N k (v) Aggregation into the k-order neighborhood feature aggregation of the fault time-frequency representation node v The feature aggregation The calculation formula is:

[0011]

[0012] The neighborhood characteristics of node v calculated by the above formula are: Where k is a fixed hyperparameter. For each fault time-frequency representation node, a corresponding multi-scale neighborhood feature can be generated. The neighborhood feature information of k scales is aggregated through the contact operation:

[0013]

[0014] in Represents the initial characteristics of the node After aggregating the features of k-order neighboring nodes, the feature matrix S of the fault time-frequency representation node will be expanded to And S k It can contain multi-scale feature information of fault feature representation nodes.

[0015] (3b) Construct the centrality encoding of the fault time-frequency representation node to save the structural information of the fault time-frequency representation node. The specific implementation process is as follows:

[0016] The degree matrix is ​​obtained based on the characteristic matrix S of the fault time-frequency representation node The degree of the fault frequency representation node is embedded into the centrality information of each node; it can be directly added to the features of the fault frequency representation node as the input of the Transformer module It is expressed as:

[0017]

[0018] in is the characteristic of the fault time-frequency representation node.

[0019] (3c) Construct the Transformer module. First, construct the Transformer encoder, which consists of two main modules: the self-attention module and the feedforward neural network. In the self-attention module, the fault time-frequency representation node features are first transformed into the fault time-frequency representation node features through linear projection. Projected to the query matrix Q, keyword matrix K and value matrix V, such that and Self-attention can be calculated by the following formula

[0020]

[0021] where d out refers to the dimension of Q, W Q , W K and W V is a learnable parameter; multi-head attention connects multiple of the above formulas together; the output of self-attention is then input into the feedforward network to form a Transformer encoder. The calculation formula of the Transformer encoder is as follows:

[0022]

[0023]

[0024] Six encoders are then stacked to form a Transformer module.

[0025] (3d) The multi-scale fault features embedded with structural information are obtained Input into the Transformer module for feature extraction.

[0026] (3e) The Transformer module is used to extract the fault features of the fault time-frequency representation node, and the features are reorganized through the fully connected layer.

[0027] 4. Input the node features after feature reorganization into the SoftMax classifier to classify the fault time-frequency representation nodes.

[0028] The beneficial effects of the present invention are as follows: a multi-scale feature aggregation module is proposed to aggregate the feature information of each neighborhood of the graph node into the feature representation of the central node, and the multi-scale node information of different neighborhoods of the node can be fused to enhance the feature representation. At the same time, the present invention uses Transformer to extract the feature information of the fused node feature information, thereby improving the utilization rate of more important node information. Experimental results show that the method of the present invention can extract more effective fault features and increase the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0030] Figure 1 is a flow chart of the method of the present invention;

[0031] Figure 2It is a three-dimensional data distribution diagram of fault characteristics;

[0032] Figure 3 is the confusion matrix of the diagnosis results. DETAILED DESCRIPTION

[0033] The specific implementation steps of the present invention are as follows

[0034] 1. Collect fault data of mechanical equipment through sensors, enhance the data through sliding windows, and obtain enhanced fault data; obtain the fault time-frequency representation node v containing both time domain and frequency domain information through wavelet transform.

[0035] 2. By calculating the distance between two fault time-frequency representation nodes, the first k fault time-frequency representation nodes of each fault time-frequency representation node v are used to construct an adjacency matrix to obtain the characteristic matrix S and adjacency matrix A of the fault time-frequency representation node.

[0036] 3. Input the feature matrix S and adjacency matrix A of the fault feature representation node into the multi-scale graph Transformer fault diagnosis model: through the multi-scale aggregation module

[0037]

[0038]

[0039] Get the feature matrix that aggregates multi-scale feature information Re-embed the structural features of the fault feature representation node

[0040]

[0041] Then As the input of the Transformer module, the Transformer module is used to extract the fault features of the fault feature representation node to extract the features of the fault node. The calculation formula of each layer of the Transformer module is as follows:

[0042]

[0043]

[0044] in

[0045] 4. Use the SoftMax classifier to classify the extracted fault features and obtain the classification results of the fault data.

[0046] The effect of the present invention can be further verified by the following experiments:

[0047] For the problem of rolling bearing fault diagnosis, the present invention uses the rolling bearing data set of Western Reserve University as an example to conduct experiments. The data used is the driving end rolling bearing fault data at a speed of 1772r / min and a sampling frequency of 12kHz. The rolling bearing type is SKF6205. The faults are all single-point damage caused by manual electrospark machining, and the damage diameter is 0.1778mm. In this experiment, the rolling bearing data is divided into ten types according to different fault types. For each type of data, we use 1024 sampling points as a sample, and a total of 1000 samples are obtained. Then, the features are divided into training sets and test sets, and input into the proposed model for training. The random experiment is repeated five times.

[0048] Figure 2 The three-dimensional data distribution of rolling bearing fault characteristics is intuitively displayed. Figure 2 It can be seen that the method of the present invention can not only effectively and accurately diagnose the fault type, but also effectively classify the fault type. Figure 3 The confusion matrix of the fault diagnosis results of the method of the present invention is shown. Figure 3 It can be seen that accurate fault diagnosis can be achieved on the driving end rolling bearing fault data at a sampling frequency of 12kHz and a rotation speed of 1772r / min. The experimental results prove that the method disclosed in the present invention is an effective rolling bearing fault diagnosis method.

Claims

1. A rolling bearing fault diagnosis method based on a multi-scale graph transformer, characterized in that The method comprises the following steps: (1) The original bearing fault data is collected through sensors, and the data is enhanced through sliding windows to obtain enhanced fault data; the fault time-frequency representation node containing both time domain and frequency domain information is obtained through continuous wavelet transform. ; (2) By calculating the distance between two fault time-frequency representation nodes, each fault time-frequency representation node Before The fault time-frequency representation nodes are used to construct the adjacency matrix and obtain the characteristic matrix of the fault time-frequency representation nodes. and the adjacency matrix ; (3) The multi-scale graph Transformer fault diagnosis model is used to extract features and learn representations of fault time-frequency representation nodes. The specific implementation method is as follows: For fault time-frequency representation nodes and ,set up for of order neighborhood, where express and The shortest path distance between , that is, when When it is 0, the 0th-order neighborhood is the feature of the node itself; the aggregation operator is used in the multi-scale feature aggregation module Will Aggregate into fault time-frequency representation nodes of Order Neighborhood Feature Aggregation , the feature aggregation The calculation formula is: ， The fault time-frequency representation node is calculated by the above formula The neighborhood characteristics are ,in is fixed as a hyperparameter; then for each fault time-frequency representation node, a corresponding multi-scale neighborhood feature can be generated; pass Operation will Aggregation of neighborhood feature information at each scale: ， in Represents the initial characteristics of the node Aggregate The characteristics of the node after the order neighborhood; then the characteristic matrix of the fault time-frequency representation node Will expand to ,and Able to contain multi-scale feature information of fault feature representation nodes; (3b) Construct the centrality encoding of the fault time-frequency representation node to save the structural information of the fault time-frequency representation node. The specific implementation process is as follows: The characteristic matrix of the node represented by the fault time-frequency Obtain the degree matrix , embed the degree of the fault frequency representation node into the centrality information for each node; it can be directly added to the feature of the fault frequency representation node as Encoder input , expressed as: ， in The characteristics of the node are represented by the fault time-frequency; Build the Transformer module: The Transformer encoder consists of two main modules: a self-attention module and a feedforward neural network. In the self-attention module, the fault time-frequency representation node features are first transformed into Projection to query matrix , Keyword Matrix Sum Matrix , so that , the self-attention can be calculated by the following formula , in means The dimension of is a learnable parameter; multi-head attention connects multiple of the above formulas together; the output of self-attention is then input into the feedforward network to form a Transformer encoder. The calculation formula of the Transformer encoder is as follows: ， Then stack six encoders to form a Transformer module; (3c) Construct the centrality encoding of the fault time-frequency representation node to save the structural information of the fault time-frequency representation node. The specific implementation process is as follows: The characteristic matrix of the node represented by the fault time-frequency Obtain the degree matrix , the degree of the fault time-frequency representation node is embedded into the centrality information of each node, which can be directly added to the features of the fault time-frequency representation node as the input of the Transformer module , expressed as: , in The characteristics of the node are represented by the fault time-frequency; (3d) The multi-scale fault features embedded with structural information are obtained Input into the Transformer module for feature extraction; (3e) Use the Transformer module to extract the fault features of the fault time-frequency representation node and reorganize the features through the fully connected layer; (4) Use the SoftMax classifier to classify the fault time-frequency representation nodes.

Images