Bearing fault diagnosis method and system based on multi-scale graph convolutional neural network

Abstract

The application belongs to the technical field of fault diagnosis, and discloses a bearing fault diagnosis method and system based on a multi-scale graph convolutional neural network. Firstly, features of original vibration signals are extracted under different scales to obtain signal information with different receptive field ranges, then mapping is performed under different scales respectively and single-scale graph convolution iteration is performed to obtain enhanced feature representation. Secondly, signal information under each single scale is spliced to obtain a cross-scale graph model, and enhanced feature representation is obtained through a graph convolution network. Finally, different features are used as priori by using Bayesian cross-fusion to obtain final result information, which is input into a full connection layer for classification to obtain a fault classification result. The application can effectively improve the rationality and accuracy of fault classification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and particularly relates to a bearing fault diagnosis method and system based on a multi-scale graph convolutional neural network. BACKGROUND

[0002] Bearings are widely used in the mechanical field, but bearing failures can cause major accidents of industrial equipment and huge maintenance costs, so bearing fault diagnosis has become one of the most important parts in modern industry, and a fault diagnosis system has become an important means to ensure the reliability of industrial processes throughout the life cycle. The fault diagnosis of rolling bearings refers to extracting the vibration signals of equipment by using an accelerometer, predicting the fault conditions in the operation of machine equipment by means of the wavelength and other information in the signals, identifying and classifying various faults, and preventing major equipment accidents.

[0003] Most of the methods for front rolling bearing fault diagnosis use deep learning to perform fault recognition, mainly using deep neural networks to extract statistical features between vibration signals to achieve classification purposes. Fault diagnosis using deep neural networks includes two key processes. One is to extract and construct graphs of vibration signals to obtain a graph model. The other is to update node information by inputting the node features of nodes and adjacency matrices to obtain fault classification output. Chenyang Li et al. proposed a rolling bearing fault diagnosis method based on weighted horizontal visibility graph and graph isomorphism network. The advantage of this method is that the horizontal visibility graph method is processed by edge weighting, which enriches the features of the graph and reduces the noise interference between sampling points. And the mapping function of the graph isomorphism network is separated, which enhances the nonlinear representation ability of the iteration process, and finally improves the classification and prediction results of fault features in non-Euclidean space.

[0004] However, this method has the following problems: First, it does not consider the information loss of feature extraction at different scales, and extracting features at a single scale can easily lead to insufficient mining of global fault context correlation (small-scale samples) and not fine enough local fault feature extraction (large-scale samples). Therefore, selecting different scales and appropriate receptive fields plays a crucial role in exploring the regularity of data signals. Second, it does not consider the construction of multi-scale graph models and the mining of potential information, and constructing graphs at a single scale ignores the correlation between signal features in multi-scale graph models, leading to insufficient mining of potential relationships between features and missing context information. Therefore, multi-scale interaction plays a crucial role in mining the potential relationships of data signals. Third, it does not consider processing the obtained signal features, making it difficult to balance the importance of different regions in the signal, resulting in insufficient accuracy of fault classification. Therefore, balancing the signal features in different regions plays a crucial role in improving the accuracy of fault classification results. SUMMARY

[0005] In view of the problems existing in the prior art, the bearing fault diagnosis method and system based on a multi-scale graph convolutional neural network are provided, signal features are extracted and mapped under multi-scale, signal features extracted under different scales are mapped through a cross-scale method, a graph convolutional neural network is used to obtain more rich feature representations, and a Bayesian model is used for cross fusion, thereby improving the accuracy of fault diagnosis.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is:

[0007] The bearing fault diagnosis method based on a multi-scale graph convolutional neural network comprises the following steps:

[0008] S1, obtaining an original vibration signal of a bearing;

[0009] S2, extracting multi-scale features: grouping the original vibration signal three times to obtain data under different scales;

[0010] S3, constructing single-scale graphs and cross-scale graphs and performing iteration: constructing single-scale graphs under three scales and performing single-scale graph convolution iteration; meanwhile, grouping data is subjected to feature extraction through a one-dimensional convolutional neural network, then the extracted features are subjected to cross-scale mapping and cross-scale graph convolution iteration;

[0011] S4, cross fusion of the features of the three single-scale graphs and the cross-scale graph obtained in step S3;

[0012] S5, outputting a fault classification result.

[0013] Preferably, in step S2, firstly, data normalization preprocessing is performed on the input signal length L of the original vibration signal data sequence X of the bearing, as shown in formula (1),

[0014] X nol =normalize(X), (1)

[0015] In the formula, X nol is a normalized time series, and normalize(X) represents different normalization methods; then the normalized data is grouped three times, and each group of data after grouping is X k , k∈1,2,3, each grouping is taken as a scale, and finally feature extraction is performed under three scales to obtain output features F k with the same dimension under three scales.

[0016] Preferably, the specific steps of extracting F k with the same dimension in step S2 are as follows:

[0017] The total dimension of the input data is d', and the data is divided into g groups in the direction of the signal dimension, so the dimension of each input vector is d' / g, i = 1, 2, …, g, and each group is independently operated in one-dimensional convolution operation, and each group is operated through one-dimensional convolution with a size of a, and the dimension of each output vector after convolution is d i i i = d' k = (a-1), and the total convolution dimension is d'-g(a-1); by adjusting the relationship between g and a, the output dimension of each scale can be ensured to be the same, and the extraction process is shown in formula (2):

[0018]

[0019] In the formula: is the parameter of CNN, X k is the signal data under different scales.

[0020] Preferably, the horizontal visibility graph method is used to construct the single scale graph model in step S3, and the affinity score between each two nodes is calculated to determine the information of the edge, specifically: three groups of vibration signals are respectively graphed by using the horizontal visibility graphing method to obtain three graph models under three single scales, and the signal data X k under different scales obtained by grouping three times in step S2 are input into the three graph models respectively, and three groups of single scale iteration generated features G k ,

[0021] Among them, the graphing process is to first convert the time series information into a graph with n nodes, and attach the corresponding data value as the node feature, and only when there is no intermediate data height between two nodes, the two nodes exist edges; on the basis of determining the nodes, the pair-wise affine of the edge feature is calculated, and the relationship is constructed by the following formula (3),

[0022]

[0023] If two computing nodes have strong data relationship, a high affinity score can be obtained; after constructing the single scale graph model of the node F and the edge R, the residual connection is applied to the original GCN to pass the aggregated information, and the single scale iteration feature is obtained, and the formula is as follows:

[0024]

[0025] In the formula, is the affinity score between two nodes and , is a training parameter, where and are obtained by backpropagation, p and q are the scales of the features, if and are nodes in the same scale, then p = q; are the parameters of the GCN, are the parameters of the residual connection, G is the iteratively generated features in three groups of single scale, G = (G1, G2, G3).

[0026] Preferably, the horizontal visibility map method is also used to construct the cross-scale graph model in step S3,

[0027] On the basis of determining the nodes, the pairwise affinities of the features of the edges are calculated, and the relationship is constructed by formula (5), if two computing nodes have strong data relationship, then a high affinity score can be obtained; after the cross-scale graph model of the node F' and the edge R' is constructed, the three groups of data X1, X2 and X3 are spliced and input into the cross-scale graph model to transmit and aggregate information, and the cross-scale iterative features G4 are obtained,

[0028]

[0029] In the formula, is the affinity score between two nodes and , is the training parameter, where and are obtained by backpropagation, j, l are the scales of the features.

[0030] Preferably, in step S4, the Bayesian mutual fusion method is used to fuse G1, G2, G3 and G4, to establish a consistent representation in different scales, to realize effective information fusion and complementarity in different scales,

[0031] The specific steps are as follows:

[0032] (1) The first step uses G1 as the prior and uses G2 as the likelihood, and then merges the information of the two features based on the Bayesian formula; the second step exchanges them, uses G2 as the prior and uses G1 as the likelihood, and then merges the information of the two features based on the Bayesian formula, denoted as G5;

[0033] (2) Similarly, the first step uses G3 as the prior and uses G4 as the likelihood, and then merges the information of the two features based on the Bayesian formula; the second step exchanges them, uses G4 as the prior and uses G3 as the likelihood, and then merges the information of the two features based on the Bayesian formula, denoted as G6;

[0034] (3) Finally, the two information G5 and G6 obtained in step (1) and step (2) are fused in the same way to obtain the final fusion feature G7.

[0035] Preferably, in step S5, the fusion feature G7 is input into a full connection layer to obtain a final fault classification, and a prediction result is obtained, as shown in formula (11).

[0036] G = softmax (FC (G7)), (11)

[0037] An NLL is used as a loss function, and its formula is shown in formula (12):

[0038] L (y G ) = -log (y G ), (12)

[0039] Wherein: G is a prediction result, FC (G7) represents that G7 is processed through a full connection layer, softmax is a softmax function, y G is a predicted label, and L (y G ) is a loss function Loss

[0040] The application also provides a bearing fault diagnosis system based on a multi-scale graph convolutional neural network, which is used to realize the bearing fault diagnosis method based on the multi-scale graph convolutional neural network as described above.

[0041] The multi-scale feature extraction module is used to obtain grouped data from the input bearing vibration signal data and extract multi-scale features.

[0042] The single-scale graph iteration module is used to construct a single-scale graph by using the multi-scale features output by the multi-scale feature extraction module and iteratively generate single-scale iteration features.

[0043] The cross-scale graph iteration module is used to construct a cross-scale graph by using the multi-scale features output by the multi-scale feature extraction module and iteratively generate cross-scale iteration features.

[0044] The cross-interfusion module is used to cross-fuse the features output by the single-scale graph iteration module and the cross-scale graph iteration module to obtain final fusion features.

[0045] The classification module is used to input the fusion features output by the cross-interfusion module into a full connection layer to obtain a final fault classification.

[0046] Compared with the prior art, the application has the following advantages:

[0047] (1) The application proposes a multi-scale method to solve the problem of incomplete signal rule exploration when processing vibration signals, better excavates the regularity of signal characteristics, and improves the rationality of fault classification. Four times of feature extraction are carried out respectively under the condition of different receptive field ranges, avoiding the problem of difficult fine extraction of local information or difficult full excavation of global information under a single scale.

[0048] (2) The application proposes a multi-scale graph convolutional neural network to solve the problem of insufficient excavation of potential relationship of signal characteristics in single scale exploration. The graph model is constructed under multiple scales and cross-scale interaction is carried out, avoiding the problem of insufficient excavation of potential relationship between signals, and better capturing the hidden information of signal characteristics under multiple scales.

[0049] (3) The application proposes a mutual fusion method based on Bayesian model to solve the problem of insufficient fault classification accuracy, better balancing the importance of data in significant and scarce areas. Different features are forced to be used as priori, effectively realizing information fusion and complementation, fully retaining complete local features on the basis of obtaining rich global information features, and improving the accuracy of fault classification under multiple scales. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0051] Figure 1 The flowchart of the application;

[0052] Figure 2 The multi-scale feature extraction module structure diagram of the application;

[0053] Figure 3 The single scale graph iteration module structure diagram of the application;

[0054] Figure 4 The cross-scale graph iteration module structure diagram of the application;

[0055] Figure 5 The cross-interfusion module structure diagram of the application. DETAILED DESCRIPTION

[0056] The application will be further described below in combination with the drawings and specific embodiments.

[0057] Embodiment 1

[0058] As Figure 1As shown, the embodiment provides a bearing fault diagnosis method based on a multi-scale graph convolutional neural network. First, the original vibration signal is grouped three times to obtain data at different scales, then graphing and single-scale graph convolution iteration are performed at three scales. In addition, the grouped data is extracted by a one-dimensional convolutional neural network, then the extracted features are cross-scale graphed and cross-scale graph convolution iteration is performed. Finally, the features obtained by the three multi-scale graph models and the features obtained by the cross-scale graph model are fused by Bayes, and the fault is classified. The following will introduce each step in detail. Figures 2-5 , and each step is introduced in detail.

[0059] The bearing fault diagnosis method based on a multi-scale graph convolutional neural network comprises the following steps:

[0060] S1, obtaining the original vibration signal of the bearing.

[0061] In this embodiment, the vibration signal of a rolling bearing is selected for fault diagnosis. The vibration data is collected using an accelerometer, and the accelerometer is fixed on the shell with a magnetic base. The bearing includes a driving end bearing and a fan end bearing, and the accelerometer is placed at the 12 o'clock position of the motor shell driving end and fan end. The sampling frequency of the driving end bearing data is 12KHZ and 48KHZ, and the sampling frequency of the fan end bearing data is 12KHZ.

[0062] S2, extracting multi-scale features: grouping the original vibration signal three times to obtain data at different scales.

[0063] In order to fully explore the law of data signal, a multi-scale feature extraction module is designed, which extracts data features at three different scales, which can greatly mine global information and local information of data signal.

[0064] In step S2, first, the input original vibration signal data of the bearing is preprocessed by data normalization. For an original data sequence X with a signal length of L, the data normalization process can be represented as formula (1); then the normalized data is grouped three times, and each group of data is X k , k e 1, 2, 3, each grouping is a scale, and finally feature extraction is performed at three scales to obtain output features F k with the same dimension at three scales.

[0065] As a preferred embodiment, the specific steps of extracting F k with the same dimension in step S2 are as follows:

[0066] The total dimension of the input data is d', and the data is divided into g groups in the direction of the signal dimension, so the dimension of each input vector is d' id' / g, i = 1, 2, …, g, each group is independently operated by one-dimensional convolution operation, each group is operated by one-dimensional convolution with a size of a, and the dimension of each output vector after convolution is d i d' i (a-1), and the total convolution dimension is d'-g(a-1); by adjusting the relationship between g and a, the output dimension of each scale can be ensured to be the same. Therefore, the data is divided into 1 group for the first time, and one-dimensional convolution operation is performed on this basis to realize the extraction of signal features; the data is divided into 2 groups for the second time, and one-dimensional convolution operation is performed on this basis; similarly, the data is divided into 4 groups for the third time, and one-dimensional convolution operation is performed on this basis. By adjusting the size of the convolution kernel in each one-dimensional convolution to 5, 32, the features extracted at three scales have the same dimension output. The extraction process is shown in formula (2):

[0067] X nol = normalize(X), (1)

[0068]

[0069] In the formula: X nol is the normalized time series, normalize(X) represents different normalization methods, is the parameter of CNN, and X k is the signal data at different scales.

[0070] S3, constructing a single scale graph and a cross-scale graph and iterating: single scale graphing and single scale graph convolution iteration are performed at three scales; meanwhile, the grouped data is subjected to feature extraction by one-dimensional convolutional neural network, and then the extracted features are subjected to cross-scale graphing and cross-scale graph convolution iteration. The following will be introduced respectively.

[0071] As a preferred embodiment, in order to study the data relationship of signal features at multiple scales, a single scale graph iteration module is designed to perform graph convolution modeling at multiple scales, which adopts a horizontal visibility graph method to construct a single scale graph model, and calculates the affinity score between each two nodes to determine the information of the edge, and the overall structure is as shown in Figure 3 .

[0072] In step S3, the horizontal visibility graph method is used to construct a single scale graph model, and the affinity score between each two nodes is calculated to determine the information of the edge. Specifically, the three groups of vibration signals are respectively subjected to graphing by the horizontal visibility graphing method to obtain three single scale graph models, and the signal data X k, k e 1, 2, 3, are input into three graph models respectively to obtain three sets of single-scale iteratively generated features G k .

[0073] wherein the process of composition is to first convert the time series information into a graph with n nodes, and attach the corresponding data value as the node feature, only when there is no intermediate data height between two nodes, the two nodes exist edges; on the basis of determining the node, the pair-wise affine of the edge feature is calculated, and the relationship is constructed through the following formula (3),

[0074]

[0075] If two computing nodes have strong data relationship, then high affinity score can be obtained; after constructing the single-scale graph model of node F and edge R, the residual connection is applied to the original GCN to pass the aggregated information, and the single-scale iterative feature is obtained, the formula is as follows:

[0076]

[0077] In the formula, is the affinity score between two nodes and , is the training parameter, wherein and are obtained by back propagation, p and q are the scales of the features, if and are nodes in the same scale, then p=q; is the parameter of GCN, is the parameter of residual connection, G is the three sets of single-scale iteratively generated features, G=(G1, G2, G3).

[0078] As a preferred embodiment, since the signal features in different scales also have potential data relationship, the embodiment designs a cross-scale graph iterative module to model the graph convolution in different scales, which is conducive to the enrichment of the context information on the graph model, and the overall structure is as shown in Figure 4 .

[0079] Since F1, F2 and F3 are data features of the same size, they can be spliced to obtain graph data of the cross-scale graph model. In step S3, the horizontal visibility composition method is also used to construct the cross-scale graph model. On the basis of determining the nodes, the pairwise affine of the features of the edges is calculated, and the relationship is constructed by formula (5). If two calculation nodes have a strong data relationship, a high affinity score can be obtained; after the cross-scale graph model of the node F' and the edge R' is constructed, the three groups of data X1, X2 and X3 are spliced and input into the cross-scale graph model to transmit aggregated information, and the cross-scale iterative feature G4 is obtained,

[0080]

[0081] In the formula, is the affinity score between two nodes and , is a training parameter, wherein and are obtained by back propagation, j, l are the scales of the features.

[0082] S4, cross-fusion the features of the three single-scale graphs and the cross-scale graph obtained in step S3.

[0083] Since the extracted signal features exist in different regions, it is very important to balance the importance between the significant regions rich in features and the background regions lacking in features. Therefore, the Bayesian cross-fusion method is used to balance in this embodiment. The cross-fusion module establishes a consistent representation of different scales, and realizes effective information fusion and complementation. The Bayesian fusion uses different signal features as a priori, effectively fuses the information of the features at different scales, and improves the accuracy of fault classification.

[0084] As shown in Figure 5 , the Bayesian cross-fusion method is used to fuse G1, G2, G3 and G4, to establish a consistent representation of different scales, to realize effective information fusion and complementation at different scales, and to improve the accuracy of learning. It integrates multi-scale information of signal features, fully retains local features on the basis of obtaining global signal features, and the specific steps are as follows:

[0085] (1) The first step uses G1 as a priori and uses G2 as a likelihood, and then combines the information of the two features based on the Bayesian formula; the second step exchanges them, uses G2 as a priori and uses G1 as a likelihood, and then combines the information of the two features based on the Bayesian formula, which is represented as G5;

[0086] (2) In the same way, the first step uses G3 as the prior and G4 as the likelihood, and then combines the information of the two features based on the Bayes formula; the second step exchanges the two, uses G4 as the prior and G3 as the likelihood, and then combines the information of the two features based on the Bayes formula, denoted as G6;

[0087] (3) Finally, the two information G5 and G6 obtained in steps (1) and (2) are fused in the same way to obtain the final fused feature G7.

[0088] Wherein, when G1 is used as the prior and G2 is used as the likelihood in step (1), the formula of the posterior probability is as shown in equation (6):

[0089]

[0090] Wherein, k∈(1, 2, 3, 4) is different scales, x is a sample, c1, c2, …, c k P(x|c1) is the probability that the sample is a fault sample under the condition that the fault type is c1, P(c2) is the probability that the fault type is c2 fault, and P2 is the posterior probability when G2 is used as the prior and G1 is used as the likelihood.

[0091] Since it is a ten-class problem, c∈[1, 10], P1 is the posterior probability when G1 is used as the prior and G2 is used as the likelihood.

[0092] Similarly, when G2 is used as the prior and G1 is used as the likelihood, the formula of the posterior probability is as shown in equation (7):

[0093]

[0094] P(x|c1) is the probability that the sample is a fault sample under the condition that the fault type is c1, P(c2) is the probability that the fault type is c2 fault, and P2 is the posterior probability when G2 is used as the prior and G1 is used as the likelihood.

[0095] Similarly, when G3 is used as the prior and G4 is used as the likelihood in step (2), the formula of the posterior probability is as follows:

[0096]

[0097] P(x|c4) is the probability that the sample is a fault sample under the condition that the fault type is c4, P(c3) is the probability that the fault type is c3 fault, and P3 is the posterior probability when G3 is used as the prior and G4 is used as the likelihood.

[0098] When G4 is used as the prior and G3 is used as the likelihood, the formula of the posterior probability is as follows:

[0099]

[0100] P(x|c3) means the probability of the sample being a fault sample in the case of the fault type being c3, P(c4) is the probability of the fault type being c4 fault, P4 is the posterior probability when G4 is used as the prior and G3 is used as the likelihood.

[0101] Similarly, we can also obtain P5 as the posterior probability when G5 is used as the prior and G6 is used as the likelihood P5; P6 is the posterior probability when G6 is used as the prior and G5 is used as the likelihood P6, the formula is as follows:

[0102]

[0103]

[0104] P(x|c6) means the probability of the sample being a fault sample in the case of the fault type being c6, P(c5) is the probability of the fault type being c5 fault; P(x|c5) means the probability of the sample being a fault sample in the case of the fault type being c5, P(c6) is the probability of the fault type being c6 fault.

[0105] Different features from four scales are integrated by the method based on Bayesian fusion. The process contains four steps. First, G1, G2 are merged to generate fusion feature G5; then, G3, G4 are merged to generate fusion feature G6; finally, G5, G6 are merged to generate fusion feature G7, the process is shown in formula (8), (9), (10):

[0106] G5=F(G1,G2)=P1+P2, (8)

[0107] G6=F(G3,G4)=P3+P4, (9)

[0108] G7=F(G5,G6)=P5+P6, (10)

[0109] Through the above steps, different features are forced to be used as prior, effectively realizing the fusion and complement of information, on the basis of obtaining rich global information features, fully retaining complete local features, better balancing the importance of data in significant and scarce areas, and improving the precision of fault classification under multi-scale. S5, output fault classification result.

[0110] The fusion feature G7 is input into the full connection layer to obtain the final fault classification, and the prediction result is obtained, the process is shown in formula (11).

[0111] G=softmax(FC(G7)), (11)

[0112] The embodiment uses NLL as a loss function, which performs well in sample classification problems. Since the larger the expected likelihood estimate is, the better, and after taking the negative sign, the smaller the negative log likelihood is, the better, the function can be well used as a loss function, and its formula is shown in formula (12):

[0113] L(y G )=-log(y G ), (12)

[0114] Wherein: G is a prediction result, FC(G7) represents that G7 is processed through a full connection layer, softmax is a softmax function, y G is a predicted label, and L(y G ) is a loss function Loss of the predicted label.

[0115] To sum up, the bearing fault diagnosis method based on the multi-scale graph convolutional neural network is mainly applied to the mechanical industry, aerospace, nuclear power field and the like. The bearing is one of the most commonly used components in the industry and the nuclear power field, and once damaged, a major accident will occur, causing personnel and property losses. Therefore, the multi-scale graph convolutional neural network is used to diagnose the vibration signal to prevent major safety accidents.

[0116] Firstly, the original vibration signal is extracted in different scales to obtain signal information with different receptive field ranges, and is respectively input to respective single-scale graph convolutional neural networks for iteration to obtain enhanced feature representation. Secondly, the signal information under each single scale is spliced to obtain a cross-scale graph model, and enhanced feature representation is obtained through a graph convolutional network. Finally, Bayesian cross-fusion forces different features to be used as a priori to obtain final result information, and the information is input to a full connection layer for classification to obtain a fault classification result.

[0117] Embodiment 2

[0118] The embodiment provides a bearing fault diagnosis system based on a multi-scale graph convolutional neural network, which is used to realize the bearing fault diagnosis method based on the multi-scale graph convolutional neural network as described in embodiment 1. The system comprises a multi-scale feature extraction module, a single-scale graph iteration module, a cross-scale graph iteration module, a cross-fusion module and a classification module.

[0119] The multi-scale feature extraction module is used to obtain grouped data from input bearing vibration signal data and extract multi-scale features.

[0120] The single-scale graph iteration module is used to construct a single-scale graph by using the multi-scale features output by the multi-scale feature extraction module and iteratively generate single-scale iteration features.

[0121] The cross-scale graph iteration module is configured to construct a cross-scale graph using the multi-scale features output by the multi-scale feature extraction module and iteratively generate cross-scale iteration features;

[0122] The cross-interfusion module is configured to cross-fuse the features output by the single-scale graph iteration module and the cross-scale graph iteration module to obtain final fused features.

[0123] The classification module is configured to input the fused features output by the cross-interfusion module into a fully connected layer to obtain a final fault classification.

[0124] The functions and implementation methods of each module can be found in the description of Embodiment 1, which will not be repeated here.

[0125] In summary, compared with the prior art, the present application has the following advantages: first, the regularity of signal features is better explored. The present application uses signal features at multiple scales to carry out fault diagnosis of rolling bearings, uses more comprehensive signal information for research, effectively mines the context information of the signal, explores the regularity of the signal features, solves the problem of incomplete exploration of signal regularity when processing vibration signals, and improves the rationality of fault classification.

[0126] Second, the potential relationship of signal features is better learned. The present application designs a multi-scale graph convolutional network to respectively research and model the signal features extracted at multiple scales, and performs weighted processing on the edges of the topological graph, effectively extracts the signal features and captures the potential information between the signal features, solves the problem of insufficient mining of feature potential relationship, and improves the accuracy of fault classification.

[0127] Third, the importance of data in significant and scarce regions is better balanced. The present application designs an inter-fusion method based on a Bayesian model to cross-interfuse features, which is conducive to balancing the importance between the significant region rich in features and the background region poor in features, enriches the learning method of the graph convolutional network, and improves the precision of fault classification at multiple scales.

[0128] Of course, the above description is not a limitation on the present application, and the present application is not limited to the above examples. Within the scope of the present application, changes, modifications, additions or substitutions made by those skilled in the art should all fall within the protection scope of the present application.

Claims

1. A bearing fault diagnosis method based on a multi-scale graph convolutional neural network, characterized in that, Comprise the following steps: S1, obtaining the original vibration signal of the bearing; S2, extract multi-scale features: group the original vibration signal three times to obtain data at three scales, and extract output features F with the same dimension k ; In step S2, F k The specific steps are as follows: The total dimension of the input data is d', and the data is divided into g groups in the direction of the signal dimension, so the dimension of each input vector is d' i = d' / g, i = 1, 2, …, g, each group is subjected to a one-dimensional convolution operation, and each group is subjected to a one-dimensional convolution with a size of a, and the dimension of each output vector after convolution is d i = d' i -(a-1), and the total convolution dimension is d'-g(a-1); by adjusting the relationship between g and a, the output dimension of each scale can be ensured to be the same, and the extraction process is shown in formula (2): In the formula: is a parameter of the CNN, X k is the signal data at different scales, is the extracted multi-scale feature; S3, constructing a single scale graph and a cross scale graph and iterating: constructing a single scale graph at three scales and iterating a single scale graph convolution; meanwhile, extracting features of the three grouped data through a one-dimensional convolutional neural network, then constructing a cross scale graph and iterating a cross scale graph convolution; S4, cross-fusing the features of the three single scale graphs and the cross scale graph obtained in step S3; S5, outputting the fault classification result.

2. The bearing fault diagnosis method based on the multi-scale graph convolutional neural network according to claim 1, characterized in that, In step S2, first, the input signal length L of the original vibration signal data sequence X of the bearing is normalized and preprocessed, as shown in formula (1): X nol = normalize(X) (1); In the formula: X nol is the normalized time series, and normalize(X) represents different normalization methods; then the normalized data is divided into three groups, and each group of data is X k , k∈1,2,3, each group is a scale, and finally feature extraction is performed under three scales to obtain output features F k with the same dimension under three scales.

3. The bearing fault diagnosis method based on the multi-scale graph convolutional neural network according to claim 2, characterized in that, The horizontal visibility graph method is used in step S3 to construct a single scale graph model, and the affinity score between each two nodes is calculated to determine the information of the edge. Specifically, the three groups of vibration signals are respectively graphed by using the horizontal visibility graph method to obtain three graph models under a single scale, and the signal data X k obtained by grouping three times in step S2 under three scales is input into the three graph models respectively to obtain three groups of features G k generated by iteration under a single scale. Wherein, the process of graphing is to convert the time series information into a graph with n nodes, and the corresponding data value is attached as the node feature. Only when there is no intermediate data height between two nodes, the two nodes have an edge; on the basis of determining the node, the pair-wise affine of the edge feature is calculated, and the relationship is constructed through formula (3) as follows: If two calculation nodes have strong data relationship, a high affinity score can be obtained; after constructing the single scale graph model of the node and the edge, the residual connection is applied to the original GCN to pass the aggregated information, and the single scale iteration feature is obtained, as follows: where is the affinity score between two nodes and , is the training parameter, where and are obtained by backpropagation, p and q are the scales of the features, if and are the nodes in the same scale, then p = q; is the parameter of GCN, is the parameter of residual connection, G is the features generated iteratively in three groups of single scale, G = (G1, G2, G3).

4. The bearing fault diagnosis method based on the multi-scale graph convolutional neural network according to claim 3, characterized in that, In step S3, the horizontal visibility graphing method is also used to construct a cross scale graph model, On the basis of determining the node, the pair-wise affine of the edge feature is calculated, and the relationship is constructed through formula (5). If two calculation nodes have strong data relationship, a high affinity score can be obtained; after constructing the cross scale graph model of the node and the edge, the three groups of data X1, X2 and X3 are spliced and input into the cross scale graph model to pass the aggregated information, and the cross scale iteration feature G4 is obtained, where is the affinity score between two nodes and , is the training parameter, where and are obtained by backpropagation, j, l are the scales of features.

5. The bearing fault diagnosis method based on the multi-scale graph convolutional neural network according to claim 4, characterized in that, In step S4, the Bayesian fusion method is used to fuse G1, G2, G3 and G4 to establish a consistent representation of different scales, realize effective information fusion and complementation under different scales; The specific steps are as follows: (1) The first step uses G1 as the prior and G2 as the likelihood, and then combines the information of the two features based on the Bayesian formula; the second step exchanges them, uses G2 as the prior and G1 as the likelihood, and then combines the information of the two features based on the Bayesian formula, denoted as G5; (2) Similarly, the first step uses G3 as the prior and G4 as the likelihood, and then combines the information of the two features based on the Bayesian formula; the second step exchanges them, uses G4 as the prior and G3 as the likelihood, and then combines the information of the two features based on the Bayesian formula, denoted as G6; (3) Finally, the two information G5 and G6 obtained in steps (1) and (2) are fused in the same way to obtain the final fusion feature G7.

6. The bearing fault diagnosis method based on the multi-scale graph convolutional neural network according to claim 5, characterized in that, In step S5, the fusion feature G7 is input into the fully connected layer to obtain the final fault classification and obtain the prediction result, as shown in formula (11); G = softmax (FC (G7)) (11); NLL is used as the loss function, and its formula is shown in formula (12): L(y G ) = -log(y G ) (12) wherein G is the prediction result, FC(G7) represents that G7 is processed by a full connection layer, softmax is a softmax function, y G is the predicted label, L(y G ) is a loss function of the predicted label.

7. A bearing fault diagnosis system based on a multi-scale graph convolutional neural network, characterized in that, The system is used for realizing the bearing fault diagnosis method based on the multi-scale graph convolutional neural network, and the system comprises a multi-scale feature extraction module, a single-scale graph iteration module, a cross-scale graph iteration module, a cross-interfusion module and a classification module. The multi-scale feature extraction module is used for obtaining grouped data from input bearing vibration signal data and extracting multi-scale features. The single-scale graph iteration module is used for constructing a single-scale graph by using the multi-scale features output by the multi-scale feature extraction module and iteratively generating single-scale iteration features. The cross-scale graph iteration module is used for constructing a cross-scale graph by using the multi-scale features output by the multi-scale feature extraction module and iteratively generating cross-scale iteration features. The cross-interfusion module is used for cross-fusing the features output by the single-scale graph iteration module and the cross-scale graph iteration module to obtain final fusion features. The classification module is used for inputting the fusion features output by the cross-interfusion module into a full connection layer to obtain final fault classification.

Images