A rolling bearing fault diagnosis method based on multi-input parallel graph convolutional neural network

Abstract

This invention discloses a semi-supervised fault diagnosis method for rolling bearings based on a multi-input parallel graph convolutional neural network. Based on the vibration acceleration signals from various sensors under normal and different fault states of the rolling bearing, this invention employs a radius-limited nearest neighbor method, using fault samples as nodes and the similarity between fault samples as edges to construct a fault sample map for each sensor. Then, using the fault sample map from each sensor as input, a multi-input parallel graph convolutional neural network semi-supervised fault diagnosis model is constructed, and the optimal model is obtained through training and parameter tuning. The fault diagnosis method proposed in this invention has high accuracy and can achieve high fault diagnosis accuracy even with a small number of labeled samples.

Description

Technical Field

[0001] This invention relates to a semi-supervised rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network, belonging to the field of rotating machinery fault diagnosis. Background Technology

[0002] Rolling bearings are the most widely used mechanical parts in rotating machinery. In the national economy and defense, the operation of equipment such as advanced rail transit equipment, aircraft generators, and industrial and agricultural equipment all rely on rolling bearings. However, prolonged operation, coupled with adverse external environments such as high temperatures, humidity, and even corrosion, makes rolling bearings highly susceptible to wear and surface damage. Therefore, research on fault diagnosis for rolling bearings is essential. It ensures the healthy operation of mechanical equipment, enables early detection of faults, and prevents problems before they occur.

[0003] Initially, fault diagnosis of rolling bearings was achieved by analyzing and processing vibration signals collected by sensors. However, with the development of data acquisition technology, the amount of fault data available has increased dramatically, rendering signal analysis-based fault diagnosis insufficient. Therefore, intelligent diagnostic techniques that organically combine machine learning models such as artificial neural networks, random forests, and support vector machines with the characteristic parameters of rolling bearings for comprehensive analysis have rapidly developed.

[0004] In recent years, with the massive scale of fault data, deep learning frameworks such as convolutional neural networks and recurrent neural networks have been applied to rolling bearing fault diagnosis. Compared to machine learning methods, these methods can automatically extract features and establish complex mapping relationships between features and fault categories. However, these methods have multi-layered network structures, requiring a large number of labeled samples to train a large number of network parameters to realize their advantages in mining and learning from massive fault data. However, in the field of fault diagnosis, the labeling of samples relies on experts with rich fault diagnosis experience and domain knowledge, and it is usually difficult to build a large number of labeled samples. With few labeled samples, the training of these methods cannot achieve a good convergence state, resulting in insufficient expressive power and poor generalization performance of the network model. Therefore, research on fault diagnosis under conditions of scarce labeled samples is a research area that is both challenging and of great application value. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network, which utilizes vibration acceleration signals returned by sensors at different positions of the rolling bearing to construct a fault diagnosis model.

[0006] This invention includes the following steps:

[0007] S1: Under normal and fault conditions of rolling bearing operation, collect vibration acceleration signals returned by sensors at different positions, and construct fault sample sets under each sensor by sampling and splicing methods;

[0008] S2: Based on the fault sample set of each sensor constructed in S1, with each fault sample as a node and the connection relationship of the fault samples as an edge, the fault sample graph of each sensor is constructed using the radius-limited nearest neighbor method.

[0009] S3: Using the fault sample maps of each sensor constructed in S2 as input, a semi-supervised fault diagnosis model based on multiple inputs is constructed by combining graph convolutional neural network layers with fully connected layers, nonlinear activation layers, etc.

[0010] S4: Based on the fault sample maps of each sensor constructed in S2, the semi-supervised fault diagnosis model based on multi-input parallel graph convolutional neural network constructed in S3 is trained, parameterized, and evaluated.

[0011] S5: Gaussian white noise is added to the input of the multi-input parallel graph convolutional neural network semi-supervised fault diagnosis model constructed in S4 to simulate the actual rolling bearing working environment, and fault diagnosis is performed based on the frequency domain characteristics of the vibration acceleration signal calculated by forward propagation.

[0012] Furthermore, step S1 constructs a fault sample set for each sensor, as detailed below:

[0013] (1) Under the conditions of normal rolling bearing and different faults, the vibration acceleration signals returned by the sensors at different positions are collected respectively.

[0014] (2) Continuously sample each vibration acceleration signal returned by each sensor under k states, once every n time moments, for a total of m samples, that is, each sample is a signal segment containing n sampling points;

[0015] (3) Vertically stitch together the samples collected by the same sensor under k states to construct a fault sample set S under the same sensor. i Therefore, the fault sample set for each sensor is S = {S...} i} i＝1,2,…,a , where a is the number of sensors.

[0016] Furthermore, step S2 constructs a fault sample map for each sensor, as detailed below:

[0017] (1) Fault sample set S under the same sensor i In this process, Fourier transform is used to extract the frequency domain features of each fault sample, and Euclidean distance is used to calculate the similarity between fault samples based on the frequency domain features.

[0018] (2) The nearest neighbor method with a limited radius is used to determine the most similar fault sample corresponding to each fault sample. That is, for any fault sample, given a distance radius, if the similarity of other fault samples to the fault sample is less than the distance radius, then these fault samples are the most similar fault samples corresponding to the fault sample.

[0019] (3) Constructing a fault sample map G under the same sensor i (V,E,F), where each fault sample is taken as a node V in the fault sample graph, its corresponding frequency domain feature is taken as the node attribute F, and the edges E of the fault sample graph are constructed based on the principle of connecting each fault sample with its most similar fault sample. Similarly, the fault sample graph under each sensor is {G}. i (V,E,F)} i＝1,2,…,a .

[0020] Furthermore, in step S3, using the fault sample images from each sensor as input, a semi-supervised fault diagnosis model based on a multi-input parallel graph convolutional neural network is constructed, as follows:

[0021] (1) Constructing graph convolutional layers

[0022]

[0023]

[0024]

[0025]

[0026] Where H (l) and H (l+1) These are the input and output of the graph convolutional layer, respectively, where A is the adjacency matrix and W is the input and output of the graph convolutional layer. (l) Let σ be the parameter matrix, and σ be the activation function. Let be the adjacency matrix after Laplace transform. For the adjacency matrix with added self-connections I, It is a degree matrix;

[0027] (2) Fault sample map G based on each sensor i Given (V, E, F), find its corresponding adjacency matrix A and feature matrix F. Using the adjacency matrix A and feature matrix F as input, construct a two-layer graph convolutional neural network to obtain the feature representation of nodes in the fault sample graph, that is, the feature representation f(S) of the fault sample under each sensor. i ), where the activation function is the ReLU function, W (0) W is the parameter matrix of the first layer. (1) The parameter matrix of the second layer

[0028]

[0029] (3) Based on the fault sample feature representation of each sensor {f(S i )} i＝1,2,…,a Construct a semi-supervised fault diagnosis model with a parallel graph convolutional neural network and multiple inputs.

[0030] Z = softmax(cat{f(S)) i )} i＝1,2,...,a )

[0031] Where cat represents the horizontal stitching of fault features extracted from each sensor, and Z represents the classification result of the fault sample.

[0032] Furthermore, step S4 involves training, parameter tuning, and evaluation of the constructed multi-input parallel graph convolutional neural network-based semi-supervised fault diagnosis model, as detailed below:

[0033] (1) Divide the fault sample set of each sensor constructed by S1 into training set, validation set and test set according to different proportions. Use the training set to train the model built by S3. Use the validation set to select the model hyperparameters (such as the number of hidden nodes, learning rate, etc.) to obtain the optimal model. Use the test set to validate the proposed model.

[0034] (2) The Adam algorithm is used to optimize the model in order to iteratively update the network weights;

[0035] (3) Use cross-entropy as the loss function to evaluate the model.

[0036]

[0037] Where y L Let C represent the set of labeled nodes, where C corresponds to the number of categories of node labels, and Y represents the number of categories. ij Z represents the true distribution of the fault sample classification. ij Predict the distribution of fault samples for classification.

[0038] Furthermore, step S5 performs fault diagnosis based on a multi-input parallel graph convolutional neural network semi-supervised fault diagnosis model, as detailed below:

[0039] (1) Under the normal rolling bearing and different fault conditions, add different levels of Gaussian white noise to simulate the noise in actual industrial applications, and then collect the vibration acceleration signals returned by sensors at different positions.

[0040] (2) The vibration acceleration signal sample is forward-propagated according to the fault diagnosis model after S4 evaluation, and the frequency domain characteristics of the vibration acceleration signal sample are calculated for fault diagnosis.

[0041] (3) The rolling bearing is subdivided into k states, including normal and different faults, to determine the accuracy of fault diagnosis in each state.

[0042] The beneficial effects of this invention are as follows:

[0043] 1. The present invention differs from other methods in that: First, the present invention uses the radius-limited nearest neighbor method to construct the fault sample map; second, the present invention constructs and trains a multi-input parallel graph convolutional neural network model based on multiple sensors to realize the identification and diagnosis of fault states.

[0044] 2. The method proposed in this invention has higher fault diagnosis accuracy compared with other methods, and can still obtain accurate fault diagnosis results even in the presence of noise interference and a small number of training experimental samples. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly described below. Referring to the accompanying drawings will provide a clearer understanding of the features and advantages of the present invention. The drawings are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort. Wherein:

[0046] Figure 1 This is a flowchart of the rolling bearing fault diagnosis method of the present invention;

[0047] Figure 2 This is a schematic diagram illustrating the construction of the fault sample set according to the present invention;

[0048] Figure 3 This is a schematic diagram illustrating the construction of the fault sample diagram of the present invention;

[0049] Figure 4 This is a diagram of the multi-input parallel graph convolutional neural network structure of the present invention. Detailed Implementation

[0050] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0052] like Figure 1 As shown, this invention uses vibration acceleration signals from various sensors under normal and different fault states of rolling bearings. It employs a radius-limited nearest neighbor method, using fault samples as nodes and the similarity between fault samples as edges, to construct a fault sample map for each sensor. Then, using the fault sample map from each sensor as input, a parallel graph convolutional neural network semi-supervised fault diagnosis model with multiple inputs is constructed. The optimal model is obtained through training and parameter tuning.

[0053] To facilitate understanding of the above technical solutions of the present invention, the following detailed explanation of the above technical solutions of the present invention will be provided through actual data processing and model building.

[0054] Example 1

[0055] A semi-supervised rolling bearing fault diagnosis method based on multi-input parallel graph convolutional neural networks includes the following steps:

[0056] S1: This invention takes the 6205-2RS JEM SKF deep groove ball bearing as an example, considering the bearing's normal operation and nine fault states: inner ring fault (F1), rolling element fault (F2), and outer ring fault at the 3 o'clock position (F3) with a fault diameter of 0.1778 mm; inner ring fault (F4), rolling element fault (F5), and outer ring fault at the 3 o'clock position (F6) with a fault diameter of 0.3556 mm; and inner ring fault (F7), rolling element fault (F8), and outer ring fault at the 3 o'clock position (F9) with a fault diameter of 0.5334 mm. Vibration acceleration signals r1(t) and r2(t) at the drive end and fan end of the rolling bearing are collected under each of the ten states, where t is the sampling time.

[0057] Since the sensor collects 12,000 vibration acceleration signals per second at a sampling frequency of 12 kHz, approximately 400 points are collected per revolution of the bearing. Therefore, the vibration acceleration signals r1(t) / r2(t) of the drive end / fan end collected when the rolling bearing is operating in 10 states are sampled once every 400 consecutive points, for a total of 100 samples per state.

[0058] like Figure 2 As shown, at the drive end / fan end, 1000 fault sample signals sampled under 10 states of the rolling bearing are vertically spliced ​​to obtain a 1000×400 drive end fault sample set S1 and a 1000×400 fan end fault sample set S2, respectively.

[0059] S2: As Figure 3 As shown, based on the fault sample sets S1 / S2 of the drive / fan end constructed in S1, Fourier transform is used to extract the frequency domain features of each fault sample, and Euclidean distance is used to calculate the similarity between the frequency domain features of the fault samples. Based on this similarity, the nearest neighbor method with a limited radius is used to determine the most similar fault sample corresponding to each fault sample. Then, each fault sample is taken as a node V in the fault sample graph, and the corresponding frequency domain feature is taken as the node attribute F. The edges E of the fault sample graph are constructed according to the principle of connecting each fault sample with its most similar fault sample, thereby constructing the fault sample graphs G1(V,E,F) / G2(V,E,F) of the drive / fan end.

[0060] S3: Construct graph convolutional layers

[0061]

[0062]

[0063]

[0064]

[0065] Where H (l) and H (l+1) These are the input and output of the graph convolutional layer, respectively, where A is the adjacency matrix and W is the input and output of the graph convolutional layer. (l) Let σ be the parameter matrix, and σ be the activation function. Let be the adjacency matrix after Laplace transform. For the adjacency matrix with added self-connections I, It is a degree matrix;

[0066] like Figure 4 As shown, the fault sample maps G1(V,E,F) / G2(V,E,F) constructed by S2 at the drive end / fan end are used as input to construct a semi-supervised fault diagnosis model based on a parallel graph convolutional neural network with multiple inputs.

[0067] First, the adjacency matrix A1 / A2 and feature matrix F1 / F2 of the fault sample graph at the drive end / fan end are obtained. Using these two matrices as input, a two-layer graph convolutional neural network is constructed to obtain the feature representations of nodes in the fault sample graph, i.e., the feature representations f(S1) / f(S2) of the fault samples at the drive end / fan end. The ReLU activation function is used. (0) / W2 (0) The parameter matrix of the first layer for the driver / fan end, W1 (1) / W2 (1) The parameter matrix for the second layer of the driver / fan end

[0068]

[0069]

[0070] Based on the feature representations f(S1) / f(S2) of the fault samples at the drive end / fan end, a graph neural network classification model with parallel graph convolutional layers with multiple inputs is constructed.

[0071] Z = softmax(cat{f(S1),f(S2)})

[0072] Where cat represents the horizontal stitching of fault features extracted from each sensor, and Z represents the classification result of the fault sample.

[0073] S4: The fault sample set for the drive / fan end constructed in S1 is divided into training, validation, and test sets according to different proportions, as shown in Table 1. The model built in S3 is trained using the training set, and the model hyperparameters are selected using the validation set. The number of hidden nodes is set to 200, the learning rate to 0.001, the number of iterations to 1000, and the radius to be limited to 1. Cross-entropy is used as the loss function to evaluate the model.

[0074]

[0075] Where y L This represents a labeled faulty sample node, where C corresponds to the number of categories of the node's labels, and Y... ij Z represents the true distribution of the fault sample classification. ij The model is classified and predicted for fault samples. Based on the principle of minimizing the loss function, the Adam algorithm is used to optimize the model by iteratively updating the network weights to obtain the optimal model.

[0076] The proposed model was validated using a test set. The fault diagnosis results were evaluated using the accuracy, precision, recall, and F1 score calculated according to the following formulas.

[0077]

[0078]

[0079]

[0080]

[0081] Where TP represents samples predicted as positive and actually positive, FN represents samples predicted as negative and actually positive, FP represents samples predicted as positive and actually negative, and TN represents samples predicted as negative and actually negative, as shown in Table 1. The method proposed in this invention can obtain accurate fault diagnosis results on labeled datasets of different sizes.

[0082] Table 1 Evaluation Indicators for Rolling Bearing Fault Diagnosis under Different Proportional Labels

[0083] training set Validation set test set Loss value accuracy Precision Recall F1_Score 0.8 0.1 0.1 0.0037 1.0000 1.0000 1.0000 1.0000 0.7 0.1 0.2 0.0046 1.0000 1.0000 1.0000 1.0000 0.6 0.1 0.3 0.0055 1.0000 1.0000 1.0000 1.0000 0.5 0.1 0.4 0.0062 1.0000 1.0000 1.0000 1.0000 0.4 0.1 0.5 0.0066 1.0000 1.0000 1.0000 1.0000 0.3 0.1 0.6 0.0071 1.0000 1.0000 1.0000 1.0000 0.2 0.1 0.7 0.0088 1.0000 1.0000 1.0000 1.0000 0.1 0.1 0.8 0.2567 0.9012 0.9012 0.9012 0.9012

[0084] With the same proportions of training set, validation set, and test set, the fault diagnosis method proposed in this invention achieves more accurate fault diagnosis results compared with convolutional neural network method (1D_CNN), graph convolutional neural network method (GCN), support vector machine (SVM) and graph semi-supervised learning method based on K-nearest neighbor graph construction (KNN_LP), as shown in Table 2.

[0085] Table 2 Evaluation Indicators for Different Methods at Given Proportions

[0086]

[0087] S5: In practical industrial applications, the working environment is extremely complex. Furthermore, noise in industrial settings is unavoidable, and vibration acceleration signals measured by vibration acceleration sensors are easily interfered with by noise. After the vibration acceleration signal is interfered with by noise, the human eye can hardly distinguish the vibration characteristics of the original vibration acceleration signal. Therefore, Gaussian white noise with σ = 0.05 is added to the fault sample set at the drive end / fan end constructed in S1 to simulate the noise in practical industrial applications. Based on the fault diagnosis model evaluated in S4, forward propagation is performed to calculate the frequency domain characteristics of 10 states of rolling bearing operation (normal and different faults) for fault diagnosis. Further subdivision is then performed to determine the accuracy of fault diagnosis in each state. The experimental results are shown in Table 3.

[0088] Table 3 Experimental results of rolling bearing fault diagnosis using the method of the present invention.

[0089]

Claims

1. A method for diagnosing rolling bearing faults based on a multi-input parallel graph convolutional neural network, characterized in that... include: S1: Under normal and fault conditions of rolling bearing operation, collect vibration acceleration signals returned by sensors at different positions; By sampling and splicing methods, a fault sample set for each sensor is constructed; S2: Based on the fault sample set of each sensor constructed in S1, with each fault sample as a node and the connection relationship of the fault samples as an edge, the fault sample graph of each sensor is constructed using the radius-limited nearest neighbor method. S3: Using the fault sample maps of each sensor constructed in S2 as input, a semi-supervised fault diagnosis model based on multiple inputs is constructed by combining a graph convolutional neural network layer with a fully connected layer and a nonlinear activation layer. S4: Based on the fault sample maps of each sensor constructed in S2, the constructed multi-input parallel graph convolutional neural network semi-supervised fault diagnosis model is trained, parameterized, and evaluated. S5: Gaussian white noise is added to the input of the multi-input parallel graph convolutional neural network semi-supervised fault diagnosis model constructed in S4 to simulate the actual rolling bearing working environment, and the frequency domain characteristics of the vibration acceleration signal are calculated based on the forward propagation to complete the fault diagnosis.

2. The rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network according to claim 1, characterized in that, Step S1 involves constructing a fault sample set for each sensor, as detailed below: S1-1: Under k conditions, including normal rolling bearing operation and various faults, collect vibration acceleration signals returned by sensors at different positions. S1-2: Continuously sample each vibration acceleration signal returned by each sensor under k states, once every n time moments, for a total of m samples, that is, each sample is a signal segment containing n sampling points; S1-3: Vertically stitch together the samples collected by the same sensor under k states to construct a fault sample set S under the same sensor. i Therefore, the fault sample set under each sensor is S = {S i } i＝1,2,…,a , where a is the number of sensors.

3. The rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network according to claim 1, characterized in that, Step S2 involves constructing a fault sample map for each sensor, as detailed below: S2-3: Fault sample set S under the same sensor i In this process, Fourier transform is used to extract the frequency domain features of each fault sample, and Euclidean distance is used to calculate the similarity between fault samples based on the frequency domain features. S2-3: The nearest neighbor method with a limited radius is used to determine the most similar fault sample corresponding to each fault sample. That is, for any fault sample, given a distance radius, if the similarity of other fault samples to the fault sample is less than the distance radius, then these fault samples are the most similar fault samples corresponding to the fault sample. S2-3: Constructing a fault sample map G under the same sensor i (V,E,F), where each fault sample is taken as a node V in the fault sample graph, the corresponding frequency domain feature is taken as the node attribute F, and the edges E of the fault sample graph are constructed based on the principle of connecting each fault sample with its most similar fault sample; similarly, the fault sample graph under each sensor is {G}. i (V,E,F)} i＝1,2,…,a , where a is the number of sensors.

4. The rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network according to claim 1, characterized in that, Step S3 uses the fault sample images from each sensor as input to construct a semi-supervised fault diagnosis model based on a multi-input parallel graph convolutional neural network, as detailed below: S3-1: Constructing a graph convolutional layer Where H (l) and H (l+1) These are the input and output of the graph convolutional layer, respectively, where A is the adjacency matrix and W is the input and output of the graph convolutional layer. (l) Let σ be the parameter matrix, and σ be the activation function. Let be the adjacency matrix after Laplace transform. For the adjacency matrix with added self-connections I, It is a degree matrix; S3-2: Fault sample map G based on each sensor i Given (V,E,F), find its corresponding adjacency matrix A and characteristic matrix F; Using the adjacency matrix A and the feature matrix F as input, a two-layer graph convolutional neural network is constructed to obtain the feature representation of nodes in the fault sample graph, that is, the feature representation f(S) of the fault sample under each sensor. i ), where the activation function is the ReLU function, W (0) W is the parameter matrix of the first layer. (1) The parameter matrix of the second layer S3-3: Fault sample feature representation based on each sensor {f(S i )} i＝1,2,…,a Construct a graph neural network classification model with parallel graph convolutional layers for multiple inputs. Z=softmax(cat{f(S i )} i＝1,2,…,a ) Where cat represents the horizontal stitching of fault features extracted from each sensor, and Z represents the classification result of the fault sample.

5. The rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network according to claim 1, characterized in that, Step S4 involves training, parameter tuning, and evaluation of the constructed multi-input parallel graph convolutional neural network-based semi-supervised fault diagnosis model, as detailed below: S4-1: Divide the fault sample set of each sensor constructed in S1 into training set, validation set and test set according to different proportions. Use the training set to train the model built in S3, use the validation set to select the model hyperparameters to obtain the optimal model, and use the test set to validate the model. S4-2: The Adam algorithm is used to optimize the model to iteratively update the network weights; S4-3: Use cross-entropy as the loss function to evaluate the model. Where y L Let C represent the set of labeled nodes, where C corresponds to the number of categories of node labels, and Y represents the set of nodes with labels. ij Z represents the true distribution of the fault sample classification. ij The classification prediction distribution for fault samples.

6. The rolling bearing fault diagnosis method based on a multi-input parallel graph convolutional neural network according to claim 1, characterized in that, Step 5 is as follows: S5-1: Under k states of normal rolling bearing and different faults, add different levels of Gaussian white noise to simulate the noise in actual industrial applications, and then collect vibration acceleration signals returned by sensors at different positions. S5-2: Vibration acceleration signal samples are forward-propagated according to the fault diagnosis model evaluated in S4, and the frequency domain characteristics of the vibration acceleration signal samples are calculated for fault diagnosis. S5-3: Subdivide the rolling bearing into k states, including normal operation and different faults, to determine the accuracy of fault diagnosis in each state.

Images