CN116358871B - Rolling bearing weak signal composite fault diagnosis method based on graph rolling network - Google Patents

Abstract

The invention discloses a rolling bearing weak signal composite fault diagnosis method of a graph rolling network, which adopts a multi-scale entropy method, signals are segmented into N sections through a sliding window, then the time domain signals of each section are subjected to fine composite multi-scale entropy feature extraction, and different weights are defined for adjacent matrixes in different forms by using a Gaussian kernel function. To achieve fault diagnosis at the graph level, a representation of the entire graph is obtained using a GCN with a pooling layer and a readout layer, upon which standard machine learning techniques can be performed. Experimental verification shows that compared with a node construction method of a time domain or a frequency domain, the fine composite multi-scale entropy considers the cross correlation of all entropy values to evaluate the dynamic conditions in fault detection, and the edge connection mode enables fault information to be transmitted and updated only among fault nodes of the same type, so that the composite fault type of the rolling bearing can be effectively identified, and the fault diagnosis efficiency is improved.

Description

Weak signal composite fault diagnosis method for rolling bearings based on graph convolutional network

Technical field

The invention belongs to the technical field of rotating machinery fault diagnosis, and relates to a rolling bearing fault diagnosis method, and in particular to a rolling bearing weak signal composite fault diagnosis method based on a graph convolution network.

Background technique

As key components in large rotating machinery equipment, rolling bearings are widely used in mechanical structures such as compressors, fans, steam turbines, turbines, generators, gas turbines, aviation generators and various electric motors. Minor failures of key components such as rolling bearings may indirectly affect system operation, trigger a series of chain reactions, cause equipment performance degradation, and then cause system-level failures. Prognostics and Health Management (PHM) aims to diagnose the health status of equipment and predict the occurrence of failures through data monitoring and analysis, thereby greatly improving the efficiency of condition maintenance. As two key components of the PHM system, intelligent diagnosis and prediction have been widely used in the monitoring of rotating machinery, such as aerospace engines, wind turbines, helicopters and high-speed trains. In mechanical equipment, compound faults are a typical fault. Machine learning algorithms are also widely used in the intelligent diagnosis of compound faults in mechanical equipment, and have been studied in depth. Intelligent diagnosis methods for compound faults can also be developed from two aspects: machine learning and deep learning.

Traditional machine learning methods mainly include: k-nearest neighbor algorithm, Bayesian classifier, support vector machine and artificial neural network, etc. For many years, domestic and foreign scholars have been exploring and applying traditional machine learning methods to conduct intelligent fault diagnosis of mechanical equipment. Jia Minping et al. combined variational mode decomposition and permutation entropy to identify composite faults of bearings through k-nearest neighbors. Sánchez et al. proposed a compound fault diagnosis method that combines k-nearest neighbor learning and random forest. By selecting high-quality features through feature sorting, they realized compound fault diagnosis of rolling bearings and gearboxes. Asr et al. proposed a composite fault diagnosis method based on non-naive Bayesian analysis, which can accurately identify single and composite faults of bearings. Tang Baoping et al. proposed a multi-fault diagnosis method for rotating machinery that combines orthogonal supervised linear local tangent space and least squares support vector machine to realize bearing fault identification. Wu Jun et al. achieved intelligent diagnosis of compound faults in rotating machinery by integrating two extreme learning machines and binary classifiers. However, in the face of industrial big data, traditional machine learning models have low model complexity, diagnostic accuracy depends on feature extraction and selection, and the effectiveness of extracted features largely depends on expert knowledge, which limits Application of intelligent diagnostic methods in industry.

In recent years, intelligent diagnosis methods with deep learning as the core have led to an upsurge of industrial innovative applications in fault diagnosis and predictive maintenance. At present, deep neural networks (Deep Neural Networks, DNNs) mainly include convolutional neural networks, autoencoders, deep belief networks, and recurrent neural networks. Sohaib et al. proposed a two-dimensional convolutional neural network composite fault diagnosis model, which can accurately change bearing composite faults under working conditions. Li Zhinong et al. proposed an intelligent fault diagnosis method using deep convolutional neural network to realize multi-fault identification of bearings. Pecht et al. proposed an unsupervised sparse feature learning mechanism to realize compound fault diagnosis of planetary gearboxes. Shen Changqing and others proposed an improved convolutional deep belief network model to achieve bearing composite fault diagnosis through multi-layer feature fusion technology. In order to overcome the uncertainty caused by manual feature extraction, selection and separation, the composite fault diagnosis method based on deep learning can achieve end-to-end learning from raw data to the health status of mechanical equipment. But most of them ignore the interdependence between data.

Graph Neural Networks (GNNs) are different from traditional deep learning methods and can mine the relationships between nodes from irregular data. It is designed to model graph data in which the relationships between nodes are reflected in the connected edges, and the edge weights reflect the strength of the relationship. Convolutions in Convolutional Neural Networks (CNNs) also have the same function. Because GNNs have good local structure modeling and node interaction capabilities in graph data, GNNs have been widely used in health monitoring of mechanical equipment. Typical representatives of graph neural networks include Graph Convolutional Networks (GCNs), Graph Recurrent Neural Networks (GRNNs), Graph Auto-Encoders (GAEs) and graph attention networks. (Graph Attention Network, GAT). GNNs can model three tasks, including: 1) node classification, where GNNs are used to obtain the representation of each node and perform node classification; 2) graph classification, where GNNs are used to obtain the representation of the entire graph and then perform graph classification; 3) Edge prediction, where GNNs are used to mine the relationships between each node and predict missing edges. Recently, GNNs are gradually used by researchers in PHM due to their ability to model the interdependencies between data and embed them into extracted features. Yu et al. used wavelet packets to decompose the vibration signals of wind turbine gearboxes and obtain graph datasets, and then implemented fault classification through the proposed fast deep GCNs. Li et al. converted the vibration signals of rolling bearings into horizontal visibility maps, and then applied GNNs to model the map data and achieve fault classification. Gao et al. proposed a rolling bearing fault diagnosis method that combines the Weighted Horizontal Visibility Graph (WHVG) and the graph Fourier transform method, which can effectively identify most of the fault impact components and has strong anti-interference ability. Li et al. used the WHVG method to convert into graph data, and the edges were weighted by the difference between sampling indicators, and proposed a graph convolution network combined with WHVG for bearing fault diagnosis. Yu et al. combined intrinsic time scale decomposition and graph signal processing to accurately and effectively identify composite faults of aeroengine rolling bearings. Yang et al. used multi-sensor data to obtain the spectrum of fault signals, constructed a label relationship graph of the adjacency matrix, and proposed a deep capsule graph convolution network to diagnose composite faults under various working conditions. Zhou et al. extracted singular values from multi-sensor vibration signal samples as sample node representations to construct graph data, and used GCNs to extract features from the graph data for feature fusion to achieve fault classification. Chen et al. proposed a GCNs fault diagnosis method that combines available measurement values with prior knowledge. They converted the pre-diagnosis results into correlation graphs and introduced weight coefficients in the GCNs model to adjust the influence of measurement values and prior knowledge, which can Implement fault diagnosis under some fault tags. In order to solve the problem that a large amount of data is unlabeled data in actual scenarios, Wang et al. proposed a graph convolution network fault diagnosis method based on vibration indicators, which verified that it has good application prospects for bearing fault diagnosis with less labeled data. . Through the literature review, it can be found that the main difference between methods based on DNNs and methods based on GNNs lies in the construction of the graph and the design of the model. Therefore, how to convert raw data into a graph structure and design a suitable GNNs network suitable for solving rolling bearing composite faults is a problem faced.

To sum up, in the mechanical power transmission system, due to the randomness, concurrency, and secondary characteristics of faults, faults occur frequently. Traditional deep learning methods mainly build models for single faults. When a compound fault occurs, the fault characteristic intensity distribution is uneven and superimposes and interferes with each other, making it difficult for traditional deep learning methods to determine which fault type it is. Whether it is traditional machine learning or deep learning, the data are related research based on Euclidean structure data, which treats all sample data as a whole for learning to achieve the purpose of classification. Algorithms based on graph convolutional networks can effectively mine the correlation information between samples that cannot be utilized by traditional deep learning methods. Therefore, it is very important to develop a rolling bearing composite fault diagnosis method based on graph convolutional network.

Contents of the invention

The purpose of this invention is to construct a fine composite multi-scale entropy graph model through the road graph topology, combine the fine composite multi-scale entropy with the graph convolution network, and provide a rolling bearing weak signal composite fault diagnosis method of the graph convolution network.

The purpose of the present invention is achieved through the following technical solutions:

A graph convolution network weak signal composite fault diagnosis method for rolling bearings, including the following steps:

Step 1: Segment the time domain signal, and then extract the fine composite multi-scale entropy feature of each segment of the signal to obtain the entropy value at multiple scales. Use the entropy sequence of each segment of the signal as a node of the road map. Establish a road map model based on fine composite multi-scale entropy; use Gaussian kernel function weighting to define different weights for different forms of adjacency matrices, and calculate the Laplacian matrix through the degree matrix and adjacency matrix as a matrix representation of the graph ;

Step 2: Collect the vibration signal data of the rolling bearing under different fault states and calculate the entropy value at multiple scales. Use the samples as vertices to construct a road map and calculate its corresponding adjacency matrix. Input the scrambled data set and adjacency matrix into In GCN, graph-level fault diagnosis is implemented.

Compared with the existing technology, the present invention has the following advantages:

1. The road map itself conforms to the fact that the vibration signal is a continuous time series signal and satisfies the regularity of the vibration signal changing with time. Therefore, the vibration and impact rules of the bearing can be obtained from the signal.

2. The GCN model based on the graph task has a higher recognition accuracy than the GCN model on the node task. The graph task uses the entire graph as a model for training, including the regularity between internal nodes, and analyzes the similarities and differences between graphs. Difference can better obtain the characteristic information of the signal within a certain interval.

3. The graph convolution network based on the road graph topology can obtain higher diagnostic accuracy with a small amount of data. Data samples constructed with multi-scale entropy can still obtain better diagnostic accuracy with a small amount of sample data. , and can accurately identify composite fault types of bearings.

4. Experimental verification shows that compared with the node construction method in the time domain or frequency domain, fine composite multi-scale entropy considers the cross-correlation of all entropy values to evaluate the dynamic conditions in fault detection. The edge connection method makes the fault information only Transferring updates between fault nodes of the same type can effectively identify composite fault types of rolling bearings and improve the efficiency of fault diagnosis.

Description of the drawings

Figure 1 is a composite fault diagnosis model for rolling bearings based on graph convolution network;

Figure 2 shows the time domain and frequency domain waveforms of the vibration signal in the 10 states of the rolling bearing;

Figure 3 shows the classification accuracy of different fault diagnosis methods (single fault diagnosis);

Figure 4 shows the confusion matrix of different fault diagnosis methods (single fault diagnosis);

Figure 5 shows the power transmission system test bench;

Figure 6 shows the classification accuracy of different fault diagnosis methods (composite fault diagnosis);

Figure 7 shows the confusion matrix of different fault diagnosis methods (composite fault diagnosis).

Detailed ways

The technical solution of the present invention will be further described below in conjunction with the accompanying drawings, but it is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention shall be covered by the present invention. within the scope of protection.

The present invention provides a rolling bearing weak signal composite fault diagnosis method of a graph convolution network. First, a road map model construction method based on fine composite multi-scale entropy is proposed. The multi-scale entropy method is used to segment the signal through a sliding window into N segments, and then perform fine composite multi-scale entropy feature extraction on the time domain signal of each segment, and apply the Gaussian kernel function to define different weights for different forms of adjacency matrices. Finally, a composite fault diagnosis model for rolling bearings based on graph convolutional networks is proposed. In order to achieve graph-level fault diagnosis, GCN with a graph pooling layer and a readout layer is used to obtain the representation of the entire graph. After obtaining the graph representation, it can be Standard machine learning techniques are performed on this representation. The specific implementation steps are as follows:

Step 1. Construct a road map model based on fine composite multi-scale entropy

Step 11. Construction of road map model based on fine composite multi-scale entropy:

Segment the time domain signal into N segments through the sliding window, l is the sample length of the original data, n is the sub-sample length, and then the time domain signal of each segment is extracted with fine composite multi-scale entropy features to obtain the entropy values at multiple scales, and the entropy sequence under each sliding window is As a node of the road map, a road map model based on fine composite multi-scale entropy is established.

Step 12: Adjacency matrix weighting method of road graph model:

Gaussian kernel function weighting is used to define different weights for different forms of adjacency matrices. The connecting edge weights under Gaussian kernel function weighting can effectively reduce the differences between vertices. The Laplacian matrix is calculated through the degree matrix and adjacency matrix. , thereby representing the graph.

Step 2: Construct a rolling bearing composite fault diagnosis model based on graph convolution network

Step 21. Collect data and calculate the adjacency matrix:

Collect vibration signal data of rolling bearings under different fault states, perform fine composite multi-scale entropy processing on them to obtain entropy values at multiple scales; use all samples as vertices to construct a road map, and use the Gaussian kernel function distance calculation formula to calculate the corresponding adjacency matrix, and divide the data set into a training set and a test set and randomly shuffle them.

Step 22. Composite fault diagnosis of rolling bearings using graph convolution network:

Input the adjacency matrix and data set in step 21 into the trained GCN network, use the forward propagation model to perform fault diagnosis on the rolling bearing test sample with unknown label, and the final fault type is passed through the output of the Softmax classifier to achieve graph level Troubleshooting.

Example:

In this embodiment, the weak signal composite fault diagnosis method of rolling bearings based on graph convolution network includes the following steps:

Step 1: Construct a road map model based on fine composite multi-scale entropy;

Step 2: Construct a rolling bearing composite fault diagnosis model based on graph convolution network.

The composite fault diagnosis model of rolling bearings based on graph convolution network proposed in this embodiment is shown in Figure 1. Using the bearing fault data in the CWRU data set, the Case Western Reserve University bearing experimental platform mainly consists of an asynchronous motor, a torsion sensor, and an AC power measurement system. It is composed of power machine and power indicator. The bearing under test is a 6205SKF deep groove ball bearing close to the driving end of the motor. The damage dimensions of the bearing inner ring, outer ring and rolling elements are 0.1778mm, 0.3556mm, 0.5334mm, the rotation speed is 1797r/min, and the sampling frequency is 12kHz. 200 groups are collected under each type of fault state, and each group contains 2048 sample points. First, _the maximum and minimum normalization method needs to _be used for normalization processing, and _the input _{signal is} divided _into Several subsamples, among which: Π is the obtained sample set, x represents the subsample, y represents the label, N represents the number of subsamples, l is the sample length of the original data, and n is the subsample length. Each subsample is assigned a corresponding label, and a sliding window of length 2048 is utilized to truncate the original signal without overlapping. The length d of each subsample is set to 2048. Then, 10 samples can be obtained. In the construction process of the road map, a graph is constructed for every 10 subsamples. The data is numbered from 0 to 9, and each subsample is connected in chronological order. After passing the original data through fine composite multi-scale entropy, the total number of samples is 50,000, including 10 types of bearing faults, and each fault contains 5,000 weighted entropy samples. Therefore, the data of its road graph structure is represented as a graph consisting of 10 nodes and 9 edges, and each node is composed of a 1×25-dimensional vector, so the total number of nodes in the sample is 2,000. 80% of each data set is randomly selected as the training set, and the remaining 20% is used as the test set. The number of iterations is 100, the initial learning rate is 0.01, and SGD is used as the optimization algorithm, where the momentum of SGD is 0.9. The batch size is 64, and each model is trained for 100 cycles for fault diagnosis. The learning rate attenuation strategy is also used to adjust the learning rate, and the weight attenuation value is initialized to 0.0005. The loss function is cross entropy for different fault locations and different The fault size is analyzed through feature extraction experiments. The specific parameters and fault characteristic frequencies of the bearings are shown in Table 1. The experimental parameter settings for different fault states are shown in Table 2. The time domain and frequency domain waveforms of 10 different fault types are shown in Figure 2.

Table 1 6205 bearing parameters and fault characteristic frequency

Table 2 Experimental parameter settings for different fault states

(1) Comparative experimental verification of single fault diagnosis of rolling bearings under different classification models

In order to verify the effectiveness of the method proposed in this invention, support vector machines (SVM), nearest neighbor classification algorithm (K-Nearest Neighbor, KNN), multilayer perceptron (MLP), CNN, long and short term are respectively used. A total of six diagnostic methods including Long Short-Term Memory (LSTM) and GCN model were used for verification. It is divided into the following three situations: In the first situation, in order to verify the traditional machine learning models SVM and KNN, fine composite multi-scale entropy is used to construct a graph model through the Gaussian kernel function and then perform SVM and KNN classification. In the second case, in order to verify the deep learning models MLP, CNN, and LSTM, fine composite multi-scale entropy is used to construct a one-dimensional data set, and the data sample is a 1×100 data structure for fault classification. In the third case, in order to verify the method proposed by the present invention, a graph model is constructed through the road map through fine composite multi-scale entropy, and the GCN model is used for fault classification. The classification accuracy of different fault diagnosis methods is shown in Figure 3.

It can be seen from Figure 3 of the classification accuracy of different fault diagnosis methods under 10 iterations that fine composite multi-scale entropy is used as the input of the graph model to achieve 100% recognition accuracy in fault diagnosis in the GCN model. Compared with SVM, KNN, MLP, CNN, and LSTM models, the iterative convergence speed is faster and the model is more stable. The fault identification accuracy of deep learning methods based on fine composite multi-scale entropy has been improved. The feature extraction method of fine composite multi-scale entropy keeps different bearing fault types fluctuating within a certain entropy value range, and the topology of the sampling road map considers the relationship between nodes and the entire bearing fault data at all scales. changes in the situation. In graph-level fault diagnosis, better results can be obtained by using the GCN model.

In order to intuitively evaluate the classification effect of the model, a confusion matrix is used for visual analysis. According to the classification of each fault category, the diagnostic accuracy of the test set of different fault diagnosis methods is shown in Table 3, and the confusion matrix of the test samples of different fault diagnosis methods is drawn as shown in Figure 4.

Table 3 Diagnostic accuracy rates of different fault diagnosis methods

It can be seen from Figure 4 and Table 3 that on the SVM model, samples with moderate rolling element faults are incorrectly identified as rolling element mild faults, and the identification accuracy of moderate rolling element faults is 51%. The moderate fault sample of the inner ring was mistakenly identified as a severe inner ring fault. The identification accuracy of the moderate fault of the inner ring was 84%, and the overall identification accuracy of the model was 93.5%. On the KNN model, samples with mild rolling element faults were incorrectly identified as moderate faults. The recognition accuracy of mild rolling element faults was 92%, and the overall recognition accuracy of the model was 99.19%. On the MLP model, samples with severe inner ring faults were incorrectly identified as normal, and the identification accuracy of severe inner ring faults was 90%. The sample with a severe outer ring fault was mistakenly identified as a minor rolling element fault. The identification accuracy of the severe outer ring fault was 90%, and the overall recognition accuracy of the model was 97%. On the CNN model, samples with moderate rolling element faults were incorrectly identified as moderate faults in the outer ring, and the identification accuracy of moderate rolling element faults was 90%. The severe fault sample of the outer ring was mistakenly identified as a severe fault of the rolling body. The identification accuracy of severe outer ring faults was 90%, and the overall identification accuracy of the model was 98%. On the LSTM model, the rolling body severe fault sample was incorrectly identified as a moderate outer ring fault, and the rolling body severe fault identification accuracy was 80%. A sample with a severe fault in the inner ring was mistakenly identified as a moderate fault in the outer ring, and the identification accuracy of a severe fault in the inner ring was 90%. A sample with a severe fault in the outer ring was mistakenly identified as a moderate fault in the outer ring, and the identification accuracy of a severe fault in the outer ring was 90%. The overall recognition accuracy of the model is 96%. The GCN model has 100% accuracy in overall and local fault identification, and has a high identification accuracy rate.

(2) Comparative experimental verification of rolling bearing composite fault diagnosis under different classification models

The XJTU Gearbox experimental platform is used, which consists of a drive motor, a controller, a planetary gearbox, a parallel axis gearbox and a brake, as shown in Figure 5. Among them, the motor type is a three-phase 3-horsepower motor, and the power supply is a three-phase AC. The two-stage planetary gearbox is composed of a two-stage planetary gear train. The first stage is a sun gear with 3 planetary gears, and the second stage is a sun gear with 4 planetary gears. The overall transmission ratio is 27:1, and the planetary gear bearings are rolling bearings. 61800, the bearing under test is installed on the planet wheel support seat of the planetary gearbox. Install two 1 acceleration sensors PCB352C04 in the X and Y directions of the planetary gearbox to collect vibration signals and use the signals in the Y direction. In the experiment, four types of bearing failure modes were prefabricated on the planetary gearbox, including rolling element failure, inner ring failure, outer ring failure and composite failure. Together with the normal state, a total of 5 vibration signals were collected. In addition, during the experiment, the motor speed was set to 1800r/min and the sampling frequency was set to 20.48kHz.

The sample selection method is consistent with the above-mentioned CWRU data set, and the original data is passed through fine composite multi-scale entropy to obtain a total sample number of 25,000, including 5 types of bearing faults, and each fault contains 5,000 entropy samples. Therefore, the data of its road graph structure is represented as a graph consisting of 10 nodes and 9 edges, and each node is composed of a 1×25-dimensional vector, so the total number of nodes in the sample is 1,000. 80% of each data set is randomly selected as the training set, and the remaining 20% is used as the test set. The number of iterations is 100. A total of six diagnostic methods including SVM, KNN, MLP, CNN, LSTM, and GCN models were used for verification. It is divided into the following three situations: In the first situation, in order to verify the effect of combining traditional machine learning methods SVM, KNN and graphical models, fine composite multi-scale entropy is used to construct a graphical model through Gaussian kernel function and then perform fault classification. In the second case, in order to verify the deep learning models MLP, CNN, and LSTM, fine composite multi-scale entropy is used to construct a one-dimensional data set, and the data sample is a 1×100 data structure for fault classification. In the third case, in order to verify the method proposed in this article, a graph model is constructed through the road map through fine composite multi-scale entropy, and the GCN model is used for fault classification. The classification accuracy of different fault diagnosis methods is shown in Figure 6.

It can be seen from Figure 6 that fine composite multi-scale entropy is used as the input of the graph model to achieve 100% recognition accuracy in fault diagnosis in the GCN model. Compared with SVM, KNN, MLP, CNN, and LSTM models, the iterative convergence speed is faster and the model is more stable. The feature extraction method of fine composite multi-scale entropy keeps different bearing fault types fluctuating within a certain entropy value range. The topological structure of the road map considers the relationship between nodes and the entire bearing fault data at all scales. Changes. In graph-level fault diagnosis, better results can be obtained by using the GCN model. In order to visually display the classification of each fault category, the confusion matrix of test samples for different fault diagnosis methods is drawn as shown in Figure 7. The diagnosis accuracy of the test set for different fault diagnosis methods is shown in Table 4.

Table 4 Diagnostic accuracy rates of different fault diagnosis methods

It can be seen from Figure 7 and Table 4 that on the SVM model, the rolling element fault samples are incorrectly identified as composite faults, and the rolling element fault identification accuracy is 98%. The outer ring fault sample was incorrectly identified as a composite fault, and the outer ring fault identification accuracy was 95%. The composite fault sample was incorrectly identified as an outer ring fault, and the composite fault identification accuracy was 94%. The overall recognition accuracy of the model is 97.38%. On the KNN model, the composite fault sample was incorrectly identified as an inner ring fault, the composite fault identification accuracy was 99%, and the overall model identification accuracy was 99.75%. On the MLP model, the rolling element fault sample was incorrectly identified as an outer ring fault, and the rolling element fault identification accuracy was 90%. The inner ring fault sample was mistakenly identified as a rolling element fault, and the inner ring fault identification accuracy was 90%. The composite fault sample was incorrectly identified as an outer ring fault. The composite fault identification accuracy was 90%, and the overall model identification accuracy was 94%. On the CNN model, the rolling element fault sample was incorrectly identified as a composite fault, the rolling element fault identification accuracy was 90%, and the overall model identification accuracy was 98%. On the LSTM model, the rolling element fault sample was incorrectly identified as a composite fault, the rolling element fault identification accuracy was 90%, and the overall model identification accuracy was 98%. The GCN model has 100% accuracy in overall and local fault identification, and has a high identification accuracy rate.

Claims (1)

1. A rolling bearing weak signal composite fault diagnosis method of a graph rolling network is characterized by comprising the following steps:

step one, segmenting a time domain signal, extracting fine composite multi-scale entropy characteristics of each segment of signal, solving entropy values under multiple scales, taking an entropy sequence of each segment of signal as a node of a road map, and establishing a road map model based on fine composite multi-scale entropy; different weight definitions are carried out on adjacent matrixes in different forms by adopting Gaussian kernel function weighting, and a Laplacian matrix is calculated through a degree matrix and the adjacent matrix and is used as matrix representation of a graph, and the method comprises the following specific steps of:

step one, constructing a road map model based on fine composite multi-scale entropy:

the time domain signal is segmented into N segments by sliding windows,l is the sample length of the original data, n is the sub-sample lengthCarrying out fine composite multi-scale entropy feature extraction on the time domain signals of each section, solving entropy values under a plurality of scales, taking the entropy sequence under each sliding window as a node of a road map, and establishing a road map model based on fine composite multi-scale entropy;

step two, a neighbor matrix weighting mode of a road map model:

different weight definitions are carried out on adjacent matrixes in different forms by adopting Gaussian kernel function weighting, the connection edge weight under the Gaussian kernel function weighting can effectively reduce the difference between vertexes, and a Laplace matrix is calculated through a degree matrix and the adjacent matrix, so that the graph is represented;

collecting vibration signal data of the rolling bearing in different fault states, calculating entropy values under multiple scales, constructing a road map by taking a sample as a vertex, calculating a corresponding adjacent matrix, inputting a disturbed data set and the adjacent matrix into the GCN, and realizing map-level fault diagnosis, wherein the method comprises the following specific steps of:

step two, collecting data and calculating an adjacency matrix:

collecting vibration signal data of the rolling bearing in different fault states, and performing fine composite multi-scale entropy processing on the vibration signal data to obtain entropy values in multiple scales; constructing a road map by taking all samples as vertexes, calculating a corresponding adjacency matrix by utilizing a Gaussian kernel function distance operation formula, dividing a data set into a training set and a testing set, and randomly disturbing the training set and the testing set;

step two, rolling bearing compound fault diagnosis of the graph rolling network:

and step two, inputting the adjacency matrix and the data set in the step two into a trained GCN network, performing fault diagnosis on the rolling bearing test sample of the unknown label by using a forward propagation model, and outputting the final fault type through a Softmax classifier to realize image level fault diagnosis.