Method for processing bearing fault sample data imbalance by using MRS-GNN model and fault classifier

Abstract

The invention discloses a method for processing bearing fault sample data imbalance by using an MRS-GNN model and a fault classifier. The method comprises the following steps: firstly, constructing a label set for all one-dimensional vibration signals of a bearing; and constructing all vibration signal samples into graph data by using a GNN graph neural network and calculating a new embedded representation of each node, and finally, expanding the number of all minority class samples by using an MRS algorithm, so that the number is equal to the number of the samples with the maximum label number. According to the method, the problem of unbalanced bearing fault sample data in bearing fault intelligent diagnosis is effectively solved, the classification accuracy of a fault classifier on all classes can be improved, and particularly the prediction capability on minority classes of faults is improved; the detection capability of a few kinds of faults is improved, high-cost faults can be prevented, accidental shutdown time and maintenance cost are reduced, and production efficiency and equipment health are guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, device and fault classifier for processing unbalanced bearing fault sample data using an MRS-GNN model. Background Art

[0002] In recent years, with the increasing speed of industrial integration, deep learning has been applied more and more widely in industry, among which data-driven classification models have been widely used in actual production and manufacturing. However, most classification models are based on balanced data for classification tasks. In actual image classification scenarios, there is a shortage of data in certain classes, which leads to imbalanced data between classes, which greatly affects the use effect of the model. Therefore, research on imbalanced data classification is of great value. In order to solve this practical problem, researchers have conducted in-depth research. For example, Chinese patent application CN117726863A discloses an imbalanced data classification model based on CGLOW global optimization; Chinese patent CN112445709B discloses a method for solving data imbalance in AFL test models through GAN.

[0003] However, for bearing fault data, vibration signal data is the most commonly used type of bearing fault diagnosis data, because bearing faults usually cause changes in bearing vibration characteristics, and vibration signal data is usually time series data collected by accelerometers or speedometers. Characteristics of this type of fault data: 1. High dimensionality: Bearing fault data is usually high-dimensional, for example, vibration signals may contain multiple channel data from multiple sensors. 2. Timing: Bearing fault data is time series data and contains time-related information. 3. Non-stationarity: Bearing fault data may be non-stationary, that is, the statistical characteristics of the data change over time.

[0004] The excellent performance of currently commonly used deep learning models is mainly attributed to the richness of available data, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and deep belief networks (DBNs). These models are trained under the assumption that a large amount of labeled data of various fault types can be collected. However, in most operating stages of bearings, their rotating machinery operates under normal conditions, and due to cost and safety considerations, the fault state only occurs for a very short time. Therefore, the amount of data in the normal working state of the bearing far exceeds the amount of data in the fault state, that is, the number of fault samples is much smaller than the number of normal samples, which leads to the problem of data imbalance and poses a challenge to the accurate delineation of decision boundaries (accurate fault classification and prediction) for deep learning models, because data imbalance seriously affects the training effect of the model, especially for the prediction accuracy of bearing faults as a minority class. Models trained on unbalanced datasets tend to be biased towards the majority class, which may lead to poor detection performance of bearing faults as a minority class. Failure to correctly identify and predict impending bearing faults may lead to equipment shutdown or accidents. In addition, in the case of existing unbalanced data, the model may also produce a high false alarm rate, that is, incorrectly labeling normal conditions as faults. This not only increases the unnecessary workload of inspection and maintenance, but may also lead to real fault alarms being ignored. Summary of the invention

[0005] The object of the present invention is to provide a method, device and fault classifier for processing bearing fault data imbalance using an MRS-GNN model to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for processing unbalanced bearing fault sample data using an MRS-GNN model comprises the following steps:

[0008] Step S1, construct the label set y∈{y1,y2,...,y k}, label all one-dimensional vibration signal samples of the bearing collected by the sensor, and the labels of vibration signal samples of the same category are the same;

[0009] Step S2, using a GNN graph neural network to construct all one-dimensional vibration signal samples into graph data G(A,X), wherein the graph data G(A,X) includes an adjacency matrix A and a feature matrix X;

[0010] Step S3, using the GraphSage algorithm of the GNN graph neural network, perform neighbor sampling on all nodes of the graph data G(A,X) to obtain the neighbor node set N(v) of each node; then use the aggregation function to obtain the aggregated embedding representation of the neighbor nodes, and then combine the self-embedded representation of each node with the aggregated embedded representation of the neighbor nodes to obtain a new embedded representation of each node;

[0011] Step S4, traverse the label set y∈{y1,y2,...,y k}, calculate the number of labels of each category of vibration signal samples, obtain the maximum number of labels, and mark the labels of other categories as minority class labels; the minority class labels are correspondingly confirmed as minority class samples;

[0012] Step S5, using the MRS algorithm, expand the number of all minority class samples so that their number is equal to the number of samples with the maximum number of labels. The specific process is:

[0013] S51, calculate the geometric center of each minority class sample, recorded as sample center Z;

[0014] S52, calculating the Euclidean distance between each minority class sample and the sample center, and calculating the average distance, which is recorded as the minority class sample radius R;

[0015] S53, randomly select M minority class samples, and calculate the vector V from the sample center Z to these M samples; obtain a synthetic vector by synthesizing all the M vectors;

[0016] S54, using normal distribution to determine the distance between the new sample and the sample center, and generating a new sample according to the following formula (5), wherein the mean of the normal distribution is R and the variance is B;

[0017]

[0018] Among them, z new Newly generated samples, z c The sample center, λ is from a normal distribution The sampling distance in ;

[0019] S55. Repeat the above steps S51-S54 until the number of all samples reaches a balance.

[0020] Preferably, the specific process of constructing the graph data G(A,X) in step S2 is:

[0021] S21, forming a sample set x∈{x1,x2,...,x n}, i = 1, 2, 3..., where each sample x iis a one-dimensional vector containing m variables (data points), each sample x i Each of them is a separate node in the graph data G; the sample set constitutes the feature matrix X of the graph data G(A,X),

[0022]

[0023] S22. By calculating the Pearson coefficient between samples, the relationship matrix between each sample and other samples, that is, the connection relationship between samples, is obtained, thereby obtaining the adjacency matrix A of the graph data G(A,X).

[0024] Preferably, the step S22 specifically includes:

[0025] S221. Calculate sample x according to the following formula 1 i and sample x j Pearson coefficient between ;

[0026]

[0027] in:

[0028] r ij is the sample x i Sample x j Pearson correlation coefficient of

[0029] x ik and x jk is the sample x i and sample x j Kth variable (data point);

[0030] and are the means of sample i and sample j respectively;

[0031] m is the number of variables (data points) in each sample;

[0032] S222, set a threshold [0,1], when the sample x i and sample x j When the Pearson coefficient between is greater than the threshold, the sample x i and sample x j The relationship between is considered to exist in the graph data as a connection relationship, that is, there are edges,

[0033]

[0034] S223, all connection relationships are formed into a set to form an adjacency matrix A,

[0035]

[0036] Preferably, the step S3 specifically includes:

[0037] S31, based on the graph data G(A, X) built in step S2, obtain the neighbor node set N of each node according to the connection relationship between the nodes (v) ;

[0038] S32. The neighbor node aggregation embedding representation of node v is calculated by the following formula (3):

[0039]

[0040] Among them, l represents the number of layers of the graph data network; N(v) represents the set of neighbor nodes of node v, is the embedding representation of the neighbor node u in the l-1 layer, and mean represents the average value;

[0041] S33, combine the embedding representation of node v itself with the aggregated embedding representation of neighboring nodes through the following formula (4) to obtain a new embedding representation of each node,

[0042]

[0043] Where: W l is the weight matrix, σ(·) is the activation function, CONCAT is the vector concatenation operation; the node v’s own embedding representation and the aggregate embedding representation of neighboring nodes The vector concatenation operation CONCAT is used to merge, and then the activation function σ(·) is used to perform nonlinear changes to obtain the new embedding representation of node v at layer l.

[0044] Preferably, the method of the present invention further comprises the following steps:

[0045] Step S6: construct an enhanced feature matrix from all vibration signal sample sets containing the newly generated samples Then, by calculating the Pearson coefficient between samples, we get the enhanced adjacency matrix Thus, we can obtain enhanced graph data

[0046]

[0047] Preferably, the method of the present invention further comprises the following steps:

[0048] Step S7: Use the GraphSage algorithm of the GNN graph neural network to enhance the graph data All nodes of the network are sampled to obtain a new embedding representation for each node; then the softmax function is used to predict the probability of the fault category to which the sample belongs, and the cross entropy loss function is used for optimization. Finally, the L=argmaxP function is used to select the fault category with the highest probability as the fault category to which the sample belongs to form a fault classifier.

[0049] Preferably, the step S7 specifically includes:

[0050] S71, assign labels to all the newly generated samples, and obtain a new label set with the initial sample set

[0051] S72, in the enhanced graph data The method in step S2 is used to obtain the embedded representation of node v The formula is as follows:

[0052]

[0053] S73, calculate the fault category label probability distribution of node v by the following formula (7):

[0054]

[0055] Among them, softmax is the activation function, which converts the output into a probability distribution, P v represents the probability distribution of the fault category label of node v, and is optimized using the cross entropy loss function, denoted by σ(·);

[0056] S74, output the classification result, and determine the label L of node v by the following formula (8): v , the label L v Corresponding to the fault category with the highest probability,

[0057]

[0058] Among them, L v is the predicted label of node v, P v,k is the probability that node v belongs to fault category k, and argmax represents the maximum value.

[0059] The present invention also discloses a device for processing unbalanced bearing fault sample data using an MRS-GNN model, comprising:

[0060] The labeling unit is used to construct the label set y∈{y1,y2,...,y k}, assign a label to each one-dimensional vibration signal sample of the bearing collected by the sensor, and the labels of vibration signal samples of the same category are the same;

[0061] A graph data construction unit, used to construct all vibration signal samples into graph data G(A,X) by using a GNN graph neural network, wherein the graph data G(A,X) includes an adjacency matrix A and a feature matrix X;

[0062] The embedding representation calculation unit is used to use the GraphSage algorithm of the GNN graph neural network to perform neighbor sampling on all nodes of the graph data G(A,X) to obtain the neighbor node set N of each node (v) ; Then use the aggregation function to get the aggregated embedding representation of the neighboring nodes, and finally combine each node’s own embedding representation with the aggregated embedding representation of the neighboring nodes to get the new embedding representation of each node;

[0063] A sample category confirmation unit is used to traverse all labels in the label set, calculate the number of labels of each category, obtain the maximum number of labels, and the labels of the remaining categories are all minority class labels; the minority class labels are correspondingly confirmed as minority class samples;

[0064] The sample quantity expansion unit of the minority class samples is used to use the MRS algorithm to expand the quantity of all the minority class samples so that their quantity is equal to the quantity of the samples with the maximum number of labels. The specific process is as follows:

[0065] S51, calculate the geometric center of each minority class sample, recorded as sample center Z;

[0066] S52, calculating the Euclidean distance between each minority class sample and the sample center, and calculating the average distance, which is recorded as the minority class sample radius R;

[0067] S53, randomly select M minority class samples, and calculate the vector V from the sample center Z to these M samples; obtain a synthetic vector by synthesizing all the M vectors;

[0068] S54, using normal distribution to determine the distance between the new sample and the sample center, and generating a new sample according to the following formula (5), wherein the mean of the normal distribution is R and the variance is B;

[0069]

[0070] Among them, z new Newly generated samples, z c The sample center, λ is from a normal distribution The sampling distance in ;

[0071] S55. Repeat the above steps S51-S54 until the number of all samples reaches a balance.

[0072] The present invention also discloses a fault classifier for processing bearing faults using an MRS-GNN model, comprising a processor and a memory, wherein an execution program is stored in the memory, and the execution program can execute the method described above.

[0073] Preferably, the memory also stores the following method:

[0074] Step S6: construct an enhanced feature matrix from all vibration signal sample sets containing the newly generated samples Then, by calculating the Pearson coefficient between samples, we get the enhanced adjacency matrix Thus, we can obtain enhanced graph data

[0075]

[0076] Step S7: Use the GraphSage algorithm of the GNN graph neural network to enhance the graph data All nodes are sampled to obtain a new embedding representation of each node; then the softmax function is used to predict the probability of the fault category to which the sample belongs, and the cross entropy loss function is used for optimization. Finally, the L=argmaxP function is used to select the fault category with the highest probability as the fault category to which the sample belongs;

[0077] Step S8, comparing the one-dimensional vibration signal of the bearing collected in real time by the sensor with the sample to determine the bearing fault category to which the one-dimensional vibration signal belongs.

[0078] The beneficial effects of the present invention are mainly reflected in: effectively solving the imbalance problem of bearing fault sample data in intelligent diagnosis, and in the future, the vibration signal collected at the industrial site will be input into the fault classifier to realize the prediction and classification of bearing faults. The present invention takes measures to solve the imbalance problem of bearing fault sample data, which can improve the classification accuracy of the fault classifier for all categories, especially improve the prediction ability of minority faults; improve the detection ability of minority faults, prevent costly faults, reduce unexpected downtime and maintenance costs, and ensure production efficiency and equipment health. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 Schematic diagram of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0080] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0081] The present invention discloses an MRS-GNN model (MRS: MeanRadius-SMOTE; GNN: Graph Neural Network), which aims to solve the problem of imbalanced sample data of rolling bearing fault categories. Sample data imbalance means that the number of samples of different categories is different. For example, when collecting vibration signals of bearing operation, the data samples of faulty conditions are much smaller than those of normal working conditions. That is, the fault samples are regarded as one category. Of course, different fault types can be divided into multiple categories, while the samples of normal working conditions are one category.

[0082] (1) Graph data generation (using GNN graph neural network to construct graph data of bearing samples)

[0083] Traditional deep learning models are often used to process image data. In contrast, the input of the GNN graph neural network is graph data G, which includes an adjacency matrix A and a feature matrix X. Therefore, the process of converting all one-dimensional vibration signal data of the bearing collected by the sensor into graph data needs to be carefully considered. The adjacency matrix is ​​used to describe the adjacency relationship between nodes and reflect the mutual connection between different faults in complex fault topologies. However, focusing only on the relationship between directly connected nodes may ignore the potential connection between different faults. Therefore, the present invention also explores the correlation between faults by analyzing the Pearson correlation coefficient. The Pearson correlation coefficient is a statistic used to evaluate the degree of linear correlation between two variables, and its value range is between -1 and 1. When the two variables x i and x j As the correlation between nodes increases, their Pearson correlation coefficient approaches 1. Therefore, by calculating the Pearson correlation coefficient between two nodes, we can reveal potential connections that may be overlooked by direct connection relationships.

[0084] Specifically, we first construct a label set y∈{y1,y2,...,y k}, label each one-dimensional vibration signal sample of the bearing collected by the sensor, and the vibration signal samples of the same category have the same label. Then, use the GNN graph neural network to construct all vibration signal samples into graph data G(A,X), which includes the adjacency matrix A and the feature matrix X.

[0085] The vibration signal samples are formed into a sample set x∈{x1,x2,...,x n}, i = 1, 2, 3..., where each sample xi is a one-dimensional vector containing m variables (data points), and each sample x i Each of them is a separate node in the graph data G; the sample set constitutes the feature matrix X of the graph data G(A,X),

[0086]

[0087] Then, by calculating the Pearson coefficient between samples, we can get the relationship matrix between each sample and other samples, that is, the connection relationship between samples, and thus get the adjacency matrix A of the graph data G(A,X).

[0088] The sample x is calculated according to the following formula (1): i and sample x j Pearson coefficient between ;

[0089]

[0090] in:

[0091] r ij is the sample x i Sample x j Pearson correlation coefficient of

[0092] x ik and x jk is the sample x i and sample x j Kth variable (data point);

[0093] and are the means of sample i and sample j respectively;

[0094] m is the number of variables (data points) in each sample;

[0095] Next, set a threshold [0,1], when the sample x i and sample x j When the Pearson coefficient between is greater than the threshold, the sample x i and sample x j The relationship between is considered to exist in the graph data as a connection relationship, that is, there are edges,

[0096]

[0097] S223, all connection relationships are formed into a set to form an adjacency matrix A,

[0098]

[0099] (2) Feature extraction (designing a feature extractor to obtain a new embedding representation for each node in the graph data)

[0100] After converting the one-dimensional vibration signal into graph data G, a direct way to generate new node representations of finite sample categories is to directly apply SMOTE and its derivative techniques on the original node feature space.

[0101] However, this approach may encounter the following challenges: (a) the original feature space of the nodes may be sparse and high-dimensional, which makes it difficult to find two similar nodes with the same label in the interpolation operation; (b) directly performing oversampling operations in the original space ignores the inherent graph topology structure in the graph data, which may result in poor quality of the synthesized nodes.

[0102] Therefore, the present invention uses a graph neural network (GNN) feature extractor to obtain feature representations of nodes by encoding node information and graph topology into a continuous embedding representation space. This embedding representation space not only captures the features of the nodes, but also encapsulates the relationships between nodes. In general, node representations should depict the categories and intra-class relationships of samples. In other words, similar nodes should be mapped to similar positions, while dissimilar nodes should be mapped to farther positions. Therefore, when interpolation operations are performed on a limited number of nodes and their nearest neighbor nodes, the resulting embedded representation is more likely to represent new samples from the same limited categories.

[0103] It is worth noting that the GNN feature extractor has a wide range of choices and can be applied to any GNN graph neural network model. For demonstration purposes, the present invention chooses the GraphSage algorithm because it can skillfully learn from various local topologies and effectively generalize to novel structures. Specifically, the method proposed in the present invention is to use the GraphSage algorithm of the GNN graph neural network to perform neighbor sampling on all nodes of the graph data G(A,X) to obtain the neighbor node set N(v) of each node; then use the aggregation function to obtain the aggregated embedding representation of the neighbor nodes, and finally combine the self-embedded representation of each node with the aggregated embedded representation of the neighbor nodes to obtain a new embedded representation of each node.

[0104] Specifically include:

[0105] S31, based on the graph data G(A,X) built in step S2, obtain the neighbor node set N(v) of each node according to the connection relationship between the nodes;

[0106] S32. The neighbor node aggregation embedding representation of node v is calculated by the following formula (3):

[0107]

[0108] Among them, l represents the number of layers of the graph data network; N(v) represents the set of neighbor nodes of node v, is the embedding representation of the neighbor node u in the l-1 layer, and mean represents the average value;

[0109] S33, combine the embedding representation of node v itself with the aggregated embedding representation of neighboring nodes through the following formula (4) to obtain a new embedding representation of each node,

[0110]

[0111] Where: W l is the weight matrix, σ(·) is the activation function, CONCAT is the vector concatenation operation; the node v’s own embedding representation and the aggregate embedding representation of neighboring nodes The vector concatenation operation CONCAT is used to merge, and then the activation function σ(·) is used to perform nonlinear changes to obtain the new embedding representation of node v at layer l. The present invention calculates a new embedded representation of each node in the graph data, which enhances the intra-class similarity and inter-class difference of the feature information of the sample nodes, and is conducive to the subsequent generation of new samples with higher quality.

[0112] (3) Generate new samples (expand the number of samples for minority classes to achieve data balance)

[0113] After obtaining the embedding representation of each node in the embedding space constructed by the GNN feature extractor, we further represent the intra-class similarity and inter-class difference between sample nodes. This enables us to oversample minority class samples. Traverse all labels in the label set, calculate the number of labels for each category, and get the maximum number of labels. The labels of the remaining categories are all minority class labels; the minority class labels are correspondingly confirmed as minority class samples.

[0114] This application selects the widely used MeanRadius-SMOTE algorithm as the main method. The specific steps are as follows:

[0115] Calculate the geometric center of each minority class sample, recorded as sample center Z;

[0116] Calculate the Euclidean distance between each minority class sample and the sample center, and calculate the average distance, recorded as the minority class sample radius R;

[0117] Randomly select M minority class samples and calculate the vector V from the sample center Z to these M samples; obtain the synthetic vector by synthesizing all M vectors;

[0118] Use normal distribution to determine the distance between the new sample and the sample center, and generate the new sample according to the following formula (5), where the mean of the normal distribution is R and the variance is B;

[0119]

[0120] Among them, z newNewly generated samples, z c The sample center, λ is from a normal distribution The sampling distance in ;

[0121] Repeat all the above steps until all sample sizes are balanced.

[0122] Using the MeanRadius-SMOTE algorithm, new minority class samples can be generated based on the embedded representation of the nodes, thereby increasing the distribution of minority class samples in the embedding space. This helps to improve the classification performance of the minority class model. At the same time, the MeanRadius-SMOTE algorithm effectively solves the class imbalance problem in a scientific and efficient way. This is achieved by ensuring sample diversity, reducing the generation of noise samples, and preserving decision boundaries. In addition, the MeanRadius-SMOTE algorithm is easy to inherit the feature information of the original sample, and the MeanRadius-SMOTE algorithm is iteratively applied to each minority class to generate new samples.

[0123] (4) Design a fault classifier (for the final classification and prediction tasks)

[0124] Construct the enhanced feature matrix from all vibration signal sample sets containing the newly generated samples Then, by calculating the Pearson coefficient between samples, we get the enhanced adjacency matrix Thus, we can obtain enhanced graph data

[0125] Next, we use the GraphSage algorithm of the GNN graph neural network to enhance the graph data. All nodes are sampled to obtain a new embedding representation of each node; the content of this part is similar to the previous article and will not be repeated here. Then the softmax function is used to predict the probability of the fault category to which the sample belongs, and the cross entropy loss function is used for optimization. Finally, the L=argmaxP function is used to select the fault category with the highest probability as the fault category to which the sample belongs, forming a fault classifier, which specifically includes the following steps:

[0126] Assign labels to the newly generated samples and get a new label set with the initial sample set

[0127] In the enhanced graph data The method in step S2 is used to obtain the embedded representation of node v The formula is as follows:

[0128]

[0129] The fault category label probability distribution of node v is calculated by the following formula (7):

[0130]

[0131] Among them, softmax is the activation function, which converts the output into a probability distribution, P v represents the probability distribution of the fault category label of node v, and is optimized using the cross entropy loss function, denoted by σ(·);

[0132] Output the classification result and determine the label L of node v through the following formula (8): v , the label L v Corresponding to the fault category with the highest probability,

[0133]

[0134] Among them, L v is the predicted label of node v, P v,k is the probability that node v belongs to fault category k, and argmax represents the maximum value.

[0135] The fault classifier formed by recording the corresponding results of the fault samples and the one-dimensional vibration signals one by one has a clear decision boundary and can accurately predict the fault type. The one-dimensional vibration signal of the bearing collected in real time by the sensor is compared with the sample to determine the bearing fault category to which the one-dimensional vibration signal belongs. In this way, when it is predicted that the vibration signal collected in real time belongs to the fault category, the fault classifier can generate a warning, truly preventing problems before they occur.

[0136] The experimental data set includes several typical bearing faults, such as inner race fault, outer race fault, and rolling element fault. Each type of fault has different damage levels to simulate different fault conditions that may occur in actual applications. At the same time, in order to simulate different operating conditions, bearing data are collected at different speeds of 600rpm, 800rpm, and 1000rpm, with a sampling rate of 50KHz.

[0137] The present invention takes this as an example and conducts experiments on four types of faults at a speed of 800 rpm using the method of the present invention, namely: normal, inner ring fault (IF), outer ring fault (OF) and ball fault (BF). Definition of imbalance rate The P value is used to measure the severity of the imbalance between categories. The smaller the P value, the more imbalanced the number of samples between categories. In this experiment, the normal class samples are the majority class, and the fault class samples are the minority class. To evaluate the classification performance, TP is the number of samples correctly classified by the model, and FN is the number of samples misclassified by the model. The experimental results are as follows:

[0138]

[0139] It can be seen that the accuracy of predicting and classifying bearing faults after balancing the fault sample data using the method of the present invention is significantly improved under different P conditions.

[0140] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for processing unbalanced bearing fault sample data using the MRS-GNN model, characterized in that: The following steps are included: Step S1: assign labels to all one-dimensional vibration signal samples of the bearing collected by the sensor to construct a label set y∈{y1,y2,...,y k }, where vibration signal samples of the same category have the same label; Step S2, using a GNN graph neural network to construct all one-dimensional vibration signal samples into graph data G(A,X), wherein the graph data G(A,X) includes an adjacency matrix A and a feature matrix X; Step S3, using the GraphSage algorithm of the GNN graph neural network, perform neighbor sampling on all nodes of the graph data G(A,X) to obtain a neighbor node set N(v) of each node; then use an aggregation function to obtain an aggregated embedding representation of the neighbor nodes, and then combine the self-embedded representation of each node with the aggregated embedded representation of the neighbor nodes to obtain a new embedded representation of each node; Step S4, traverse the label set y∈{y1,y2,...,y k }, calculate the number of labels of each category of vibration signal samples, obtain the maximum number of labels, and mark the labels of other categories as minority class labels; the minority class labels are correspondingly confirmed as minority class samples; Step S5, using the MRS algorithm, expand the number of all minority class samples so that their number is equal to the number of samples with the maximum number of labels. The specific process is: S51, calculate the geometric center of each minority class sample, recorded as sample center Z; S52, calculating the Euclidean distance between each minority class sample and the sample center, and calculating the average distance, which is recorded as the minority class sample radius R; S53, randomly select M minority class samples, and calculate the vector V from the sample center Z to these M samples; obtain a synthetic vector by synthesizing all the M vectors; S54, using normal distribution to determine the distance between the new sample and the sample center, and generating a new sample according to the following formula (5), wherein the mean of the normal distribution is R and the variance is B; Among them, z new Newly generated samples, z c The sample center, λ is from a normal distribution The sampling distance in ; S55. Repeat the above steps S51-S54 until the number of all samples reaches a balance.

2. The method according to claim 1, characterized in that The specific process of constructing the graph data G(A,X) in step S2 is: S21, forming a sample set x∈{x1,x2,...,x n }, i = 1, 2, 3..., where each sample x i is a one-dimensional vector containing m variables, each sample x i Each of them is a separate node in the graph data G; the sample set constitutes the feature matrix X of the graph data G(A,X), S22, by calculating each sample x i With other samples x j The Pearson coefficient between them is used to obtain the relationship matrix between the two, that is, the connection relationship between samples, so as to obtain the adjacency matrix A of the graph data G(A,X).

3. The method according to claim 2, characterized in that The step S22 specifically includes S221. Calculate sample x according to the following formula 1 i and sample x j Pearson coefficient between ; in: r ij is the sample x i Sample x j Pearson correlation coefficient of x ik and x jk is the sample x i and sample x j Kth variable; and are the means of sample i and sample j respectively; m is the number of variables per sample; S222, set a threshold [0,1], when the sample x i and sample x j When the Pearson coefficient between is greater than the threshold, the sample x i and sample x j The relationship between is considered to exist in the graph data as a connection relationship, that is, there are edges, S223, all connection relationships are formed into a set to form an adjacency matrix A, 4. The method according to claim 1, characterized in that: The step S3 specifically includes: S31, based on the graph data G(A,X) that has been built in step S2, according to the connection relationship between the nodes of the graph data, obtain the neighbor node set N(v) of each node; S32. The neighbor node aggregation embedding representation of node v is calculated by the following formula (3): Among them, l represents the number of layers of the graph data network; N(v) represents the set of neighbor nodes of node v, is the embedding representation of the neighbor node u at layer l-1, and mean represents the average value; S33. Combine the embedding representation of node v itself with the aggregate embedding representation of neighboring nodes through the following formula (4) to obtain a new embedding representation for each node: Where: W l is the weight matrix, σ(·) is the activation function, CONCAT is the vector concatenation operation; the node v’s own embedding representation and the aggregate embedding representation of neighboring nodes The vector concatenation operation CONCAT is used to merge, and then the activation function σ(·) is used to perform nonlinear changes to obtain the new embedding representation of node v at layer l.

5. The method according to claim 1, characterized in that It also includes the following steps, Step S6: construct an enhanced feature matrix from all vibration signal sample sets containing the newly generated samples Then, by calculating the Pearson coefficient between samples, we get the enhanced adjacency matrix Thus, we can obtain enhanced graph data 6. The method according to claim 5, characterized in that It also includes the following steps, Step S7: Use the GraphSage algorithm of the GNN graph neural network to enhance the graph data All nodes of the network are sampled to obtain a new embedding representation for each node; then the softmax function is used to predict the probability of the fault category to which the sample belongs, and the cross entropy loss function is used for optimization. Finally, the L=argmaxP function is used to select the fault category with the highest probability as the fault category to which the sample belongs to form a fault classifier.

7. According to the method of claim 6, the step S7 specifically comprises: S71, assign labels to all the newly generated samples, and obtain a new label set with the initial sample set S72, in the enhanced graph data The method in step S2 is used to obtain the embedded representation of node v The formula is as follows: S73, calculate the fault category label probability distribution of node v by the following formula (7): in, Softmax is an activation function that converts the output into a probability distribution, P v represents the probability distribution of the fault category label of node v, and is optimized using the cross entropy loss function, denoted by σ(·); S74, output the classification result, and determine the label L of node v by the following formula (8): v , the label L v Corresponding to the fault category with the highest probability, Among them, L v is the predicted label of node v, P v,k is the probability that node v belongs to fault category k, and argmax represents the maximum value.

8. A device for processing unbalanced bearing fault sample data using an MRS-GNN model, characterized in that: include: The labeling unit is used to label all the one-dimensional vibration signal samples of the bearing collected by the sensor, and construct a label set y∈{y1,y2,...,y k }, where vibration signal samples of the same category have the same label; A graph data construction unit, used to construct all one-dimensional vibration signal samples into graph data G(A,X) by using a GNN graph neural network, wherein the graph data G(A,X) includes an adjacency matrix A and a feature matrix X; The embedding representation calculation unit is used to perform neighbor sampling on all nodes of the graph data G(A,X) by using the GraphSage algorithm of the GNN graph neural network to obtain a neighbor node set N(v) of each node; then, an aggregation function is used to obtain an aggregated embedding representation of the neighbor nodes, and then the self-embedding representation of each node is combined with the aggregated embedding representation of the neighbor nodes to obtain a new embedding representation of each node; The sample category confirmation unit is used to traverse the label set y∈{y1,y2,...,y k }, calculate the number of labels of each category, and obtain the maximum number of labels. The labels of the remaining categories are all minority labels; the minority labels are correspondingly confirmed as minority samples; The sample quantity expansion unit of the minority class samples is used to use the MRS algorithm to expand the quantity of all the minority class samples so that their quantity is equal to the quantity of samples with the maximum number of labels. The specific process is as follows: S51, calculate the geometric center of each minority class sample, recorded as sample center Z; S52, calculating the Euclidean distance between each minority class sample and the sample center, and calculating the average distance, which is recorded as the minority class sample radius R; S53, randomly select M minority class samples, and calculate the vector V from the sample center Z to these M samples; obtain a synthetic vector by synthesizing all the M vectors; S54, using normal distribution to determine the distance between the new sample and the sample center, and generating a new sample according to the following formula (5), wherein the mean of the normal distribution is R and the variance is B; Among them, z new Newly generated samples, z c The sample center, λ is from a normal distribution The sampling distance in ; S55. Repeat the above steps S51-S54 until the number of all samples reaches a balance.

9. A fault classifier for processing bearing faults using the MRS-GNN model, characterized in that: The method comprises a processor and a memory, wherein an execution program is stored in the memory, and the execution program can execute any method according to claims 1-7.

10. The fault classifier according to claim 9, characterized in that: The execution program stored in the memory may also execute the following method: Step S6: construct an enhanced feature matrix from all vibration signal sample sets containing the newly generated samples Then, by calculating the Pearson coefficient between samples, we get the enhanced adjacency matrix Thus, we can obtain enhanced graph data Step S7: Use the GraphSage algorithm of the GNN graph neural network to enhance the graph data All nodes are sampled to obtain a new embedding representation of each node; then the softmax function is used to predict the probability of the fault category to which the sample belongs, and the cross entropy loss function is used for optimization. Finally, the L=argmaxP function is used to select the fault category with the highest probability as the fault category to which the sample belongs; Step S8, comparing the one-dimensional vibration signal of the bearing collected in real time by the sensor with the sample to determine the fault category to which the one-dimensional vibration signal belongs.

Images