Industrial Robot Fault Diagnosis System and Method Based on Graph Convolutional Neural Network

Abstract

The present invention relates to the field of industrial robot fault detection, and particularly to an industrial robot fault diagnosis system and method based on a graph convolutional neural network. The following technical solutions are adopted: It includes a data acquisition module, a node construction module, and a graph structure fault detection module; the data acquisition module is used to collect fault signals and perform preprocessing; the node construction module is used to construct multiple nodes according to the fault signals; the graph structure fault detection module is used to form an adjacency matrix based on the nodes, perform graph convolution on the adjacency matrix, and train and learn to output a fault classification result. The beneficial effects are as follows: While considering the relationship between nodes, it also considers the factors of the nodes themselves, making the adjacency matrix more adaptable. At the same time, the peak change degree matrix is used to be added to the degree matrix to process the initial adjacency matrix, which can effectively improve the practicality of the adjacency matrix and improve the fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to the field of industrial robot fault detection, and particularly to an industrial robot fault diagnosis system and method based on a graph convolutional neural network. Background Art

[0002] In the fault diagnosis of industrial robots, the actual operation conditions of robots vary greatly, so the collected fault data also has different interferences, such as noise interference, robot interference, etc. Among them, machine interference mainly includes problems such as data loss and asymmetry, resulting in irregular time series. Most conventional fault diagnosis algorithms are based on expert experience to set thresholds for judgment, and the influence of human factors is relatively large. And general deep neural networks are only applicable to solving fault diagnosis with a fixed input size format.

[0003] Currently, for the situation of irregular data, most use statistical-based detection methods. For example, intercept all irregular data, intercept one cycle of each data, and then compare the similarity of the data. The indicators include: root mean square error, mean absolute error, correlation coefficient, etc. By calculating the indicators of two groups of data to judge the similarity between the two groups of data. If it is within the defined error range, it means that the sample is a normal sample. If it exceeds the defined error range, it means that the sample is a fault sample. There is also a method of directly comparing the similarity of two groups of irregular data, such as using the independent sample t-test distribution. This method is applicable to comparing the similarity degree of two groups of irregular data under the condition of consistent distribution. The indicator P is used to define faults and normal samples. The above two fault detection methods have problems such as low recognition efficiency and low diagnostic accuracy for fault samples. Summary of the Invention

[0004] The purpose of the present invention is to provide an industrial robot fault diagnosis system and a fault diagnosis method based on a graph convolutional neural network, which can achieve efficient processing of irregular data and effectively improve the accuracy and precision of fault diagnosis.

[0005] To achieve the above object, the present invention adopts the following technical solution: An industrial robot fault diagnosis system based on a graph convolutional neural network, comprising a data acquisition module, a node construction module, and a graph structure fault detection module; the data acquisition module is used to collect the fault signals of a faulty robot and preprocess the fault signals; the node construction module is connected to the data acquisition module and is used to take peak points at different scales with different-sized windows of the fault signals obtained by the data acquisition module as multiple nodes; the graph structure fault detection module includes a node connection module for forming an adjacency matrix according to the nodes constructed by the node construction module, a graph convolution processing module for performing graph convolution operations on each node in the adjacency matrix, and a hidden layer function module for training and learning to output a fault classification result.

[0006] An industrial robot fault diagnosis method based on a graph convolutional neural network includes the following steps:

[0007] S01. First, collect and preprocess the fault signals of the faulty robot, and segment the fault signal data for fault detection from them.

[0008] S02. Build nodes, intercept the window of the fault signal data in S01, and use windows of different sizes to obtain peak points at different scales for the intercepted data. The obtained peak points are the nodes, and the peak values corresponding to the peak points are the eigenvalues of the nodes.

[0009] S03. Connect the nodes obtained in S02 to construct an initial adjacency matrix A, and use the set degree matrix D to normalize the initial adjacency matrix A to obtain the final adjacency matrix A * , and the calculation formula for normalization processing is: .

[0010] S04. After performing graph convolution calculations 2 to 3 times on the final adjacency matrix A in S03 * , the calculation formula for each graph convolution is σ A * HW] , where σ[ ] is the graph convolution activation function, H is the overall feature of the final adjacency matrix A * , W is the built-in parameter of the final adjacency matrix. After each graph convolution calculation, the can be updated. After the update, perform graph convolution calculation again, repeat this 2 to 3 times, and then perform global average pooling on the overall feature H after 2 to 3 graph convolution calculations to obtain the embedding vector of the graph.

[0011] S05. Input the embedding vector of the graph obtained in S04 into the hidden layer function for training and learning to output the fault classification result.

[0012] Specifically, in S03, when connecting the nodes obtained in S02 to construct the initial adjacency matrix A, the K-means clustering algorithm is specifically used for calculation and connection.

[0013] Specifically, in S03, before using the set degree matrix D to normalize the initial adjacency matrix A, first add an identity matrix with trainable parameters to the initial adjacency matrix A for processing to obtain the processed initial adjacency matrix , where is the trainable parameter, and its initial value is 1. For each graph convolution calculation, Update, where I is the identity matrix; the calculation formula for the corresponding normalization process is: .

[0014] Specifically, in S03, when using the set degree matrix D to perform normalization processing on the initial adjacency matrix A, the peak change degree matrix T is also added for processing. The peak change degree matrix T is calculated based on all nodes in the initial adjacency matrix A. After adding the peak change degree matrix, the formula for normalization processing is: .

[0015] The beneficial effects of the present invention are as follows: By constructing nodes through wavelet transform of fault signals and connecting the nodes to obtain the initial adjacency matrix, while adding trainable parameters and improving the degree matrix to normalize the initial adjacency matrix, this takes into account both node relationships and the factors of the nodes themselves, making the adjacency matrix more adaptable. At the same time, using the peak change degree matrix and adding it to the degree matrix to process the initial adjacency matrix can effectively improve the practicality of the adjacency matrix and improve the accuracy of fault diagnosis. Description of the Drawings

[0016] Appendix Figure 1 It is the connection schematic diagram of the industrial robot fault diagnosis system based on the graph convolutional neural network in the embodiment. Specific Embodiments

[0017] Example 1, referring to Figure 1 , an industrial robot fault diagnosis system based on the graph convolutional neural network, including a data acquisition module, a node construction module, and a graph structure fault detection module; the data acquisition module is used to collect the fault signals of the faulty robot and preprocess the fault signals; the node construction module is connected to the data acquisition module and is used to take the peak points at different scales with different window sizes of the fault signals obtained by the data acquisition module as multiple nodes; the graph structure fault detection module includes a node connection module for forming an adjacency matrix according to the nodes constructed by the node construction module, a graph convolution processing module for performing graph convolution operations on each node in the adjacency matrix, and a hidden layer function module for training and learning to output the fault classification result.

[0018] Based on the above fault diagnosis system, this embodiment also provides an industrial robot fault diagnosis method based on the graph convolutional neural network, including the following steps:

[0019] S01. First, collect the fault signals of the faulty robot and preprocess them, and segment out the fault signal data for fault detection from them.

[0020] S02. Construct nodes, intercept the fault signal data in S01 with a window, and use windows of different sizes to obtain peak points at different scales for the intercepted data. The obtained peak points are the nodes, and the peaks corresponding to the peak points are the characteristic values of the nodes.

[0021] S03. Connect the nodes obtained in S02 to construct an initial adjacency matrix A, and use the set degree matrix D to normalize the initial adjacency matrix A to obtain the final adjacency matrix A * , and the calculation formula for normalization is: .

[0022] S04. After performing graph convolution calculations 2 to 3 times on the final adjacency matrix A in S03 * , the calculation formula for each graph convolution is σ A * HW] , where σ[ ] is the graph convolution activation function, H is the overall feature of the final adjacency matrix A * , W is the built-in parameter of the final adjacency matrix. After each graph convolution calculation, the can be updated. After the update, the graph convolution calculation is performed again. Repeat this 2 to 3 times, and then perform global average pooling on the overall feature H after performing 2 to 3 graph convolution calculations to obtain the embedding vector of the graph.

[0023] S05. Input the embedding vector of the graph obtained in S04 into the hidden layer function for training and learning to output the fault classification result.

[0024] In S03, when connecting the nodes obtained in S02 to construct the initial adjacency matrix A, the K-means clustering algorithm is specifically used for calculation and connection.

[0025] In a further embodiment, in S03, before using the set degree matrix D to normalize the initial adjacency matrix A, a unit matrix with trainable parameters is added to the initial adjacency matrix A for processing to obtain the processed initial adjacency matrix , where is the trainable parameter, and its initial value is 1. Each time a graph convolution calculation is performed, will be updated. I is the unit matrix; the corresponding calculation formula for normalization is: . By adding the trainable parameter and the unit matrix, while considering the node relationship, the factors of the nodes themselves are also considered, making the adjacency matrix more adaptable.

[0026] In a further embodiment, in S03, when the initial adjacency matrix A is normalized using the set degree matrix D, the peak change degree matrix T is also added for processing. The peak change degree matrix T is calculated based on all nodes in the initial adjacency matrix A. After adding the peak change degree matrix, the formula for normalization processing is: . By adding the peak change degree matrix to improve the degree matrix, the characteristics of the adjacency matrix can be made clearer and the accuracy of fault diagnosis can be improved.

[0027] Of course, the above is only a preferred embodiment of the present invention, and it does not limit the scope of use of the present invention. Therefore, all equivalent changes made on the principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An industrial robot fault diagnosis method based on a graph convolutional neural network, characterized in that, It includes the following steps: S01. First, collect and preprocess the fault signals of the faulty robot, and segment the fault signal data for fault detection from them; S02. Construct nodes, intercept the window of the fault signal data in S01, and obtain the peak points at different scales by using windows of different sizes for the intercepted data. The obtained peak points are the nodes, and the corresponding peaks of the peak points are the eigenvalues of the nodes; S03. Connect the nodes obtained in S02 to construct an initial adjacency matrix A, and use the set degree matrix D to normalize the initial adjacency matrix A to obtain the final adjacency matrix A * , and at the same time, a peak change degree matrix T is added for processing. The peak change degree matrix T is calculated based on all the nodes in the initial adjacency matrix A. After adding the peak change degree matrix, the normalization formula is: ; among them, when connecting the nodes obtained in S02 to construct the initial adjacency matrix A, the K-means clustering algorithm is specifically used for calculating the connections; S04. After performing graph convolution calculations on the final adjacency matrix A in S03 * 2 to 3 times, the calculation formula for each graph convolution is , where is the graph convolution activation function, H is the overall feature of the final adjacency matrix A * , W is the built-in parameter of the final adjacency matrix. After each graph convolution calculation, the can be updated. After the update, the graph convolution calculation is performed again. Repeat this 2 to 3 times, and then perform global average pooling on the overall feature H after 2 to 3 times of graph convolution calculations to obtain the embedding vector of the graph; S05. Input the embedding vector of the graph obtained in S04 into the hidden layer function for training and learning to output the fault classification result.

2. The industrial robot fault diagnosis method based on a graph convolutional neural network according to claim 1, wherein: In S03, before using the set degree matrix D to normalize the initial adjacency matrix A, the initial adjacency matrix A is first processed by adding an identity matrix with trainable parameters to obtain the processed initial adjacency matrix , where is a trainable parameter with an initial value of 1. Each time the graph convolution is calculated, is updated, and I is the identity matrix; the calculation formula for the corresponding normalization process is: .

3. A system for implementing the industrial robot fault diagnosis method based on a graph convolutional neural network according to any one of claims 1-2, characterized in that: It includes a data acquisition module, a node construction module, and a graph structure fault detection module; the data acquisition module is used to collect the fault signals of the faulty robot and preprocess the fault signals; the node construction module is connected to the data acquisition module and is used to take the peak points at different scales with windows of different sizes for the fault signals obtained by the data acquisition module as multiple nodes; the graph structure fault detection module includes a node connection module for forming an adjacency matrix according to the nodes constructed by the node construction module, a graph convolution processing module for performing graph convolution operations on each node in the adjacency matrix, and a hidden layer function module for training and learning to output the fault classification result.

Images