CN116662848A - Rolling bearing fault diagnosis method based on WOA-VMD and GAT - Google Patents

Abstract

The invention discloses a rolling bearing fault diagnosis method based on WOA-VMD and GAT, which comprises the following steps: step one, carrying out self-adaptive determination on the number k of parameter modes of VMD decomposition and penalty parameters alpha through a WOA optimization algorithm, so as to carry out VMD decomposition on an original signal, and screening IMF components with large correlation in the decomposed signal by adopting Pearson correlation analysis so as to reconstruct the signal and finish signal noise reduction; and secondly, combining the Attention and graph convolution operation to construct a graph Attention neural network rolling bearing fault diagnosis model, and distributing more specific gravity to the value information for optimizing the information collection stage of the constructed graph and improving the fault diagnosis accuracy of the model. Experimental verification shows that compared with a MLP, attention model and a GCN model, the method has the advantages of higher convergence rate, higher diagnosis precision and lower loss value.

Description

Rolling bearing fault diagnosis method based on WOA-VMD and GAT

Technical Field

The invention belongs to the technical field of fault diagnosis of rotating machinery, relates to a rolling bearing fault diagnosis method, and in particular relates to a rolling bearing fault diagnosis method based on Whale Optimization Algorithm (WOA) -Variation Modal Decomposition (VMD) and graphic ideographic neural network (GAT).

Background

The current internet rapidly develops, and the traditional manufacturing industry also follows the step of the time, and develops the past post-maintenance, regular maintenance, optionally maintenance and the like to active fault prediction and maintenance. When a rolling bearing malfunctions in a machine, not only is the bearing damaged, but also abnormal vibration due to the malfunction is transmitted to parts assembled with the rolling bearing and extends outwards continuously, thereby affecting the whole machine.

The working environment of the rolling bearing is complex, the external interference factors in the signal acquisition process are many, but the current rolling bearing fault diagnosis technology and state monitoring means can not exclude the interference of noise, and most of the commonly acquired vibration signals are mixed with a great amount of noise, so that great trouble is caused to the performance of mechanical equipment and the equipment health prediction. The quality of the fault diagnosis effect depends on the quality of the noise reduction effect to a certain extent, so that the noise reduction method with good effect is selected, and the screening of noise and other useless parts in the acquired signals has great engineering application significance. A common signal noise reduction method is a variational modal decomposition (EMD), wavelet transform (wavelet analysis), empirical modal decomposition (Empirical Mode Decomposition), and the like. Where EMD signal decomposition is typically accomplished using a time feature scale that the acquired data itself has. Chen Long et al address the problem of noise reduction by applying EMD to process the original signal containing the interfering signal and further performing feature extraction processing to reconstruct the extracted time domain features and frequency domain features into a new rolling bearing fault feature set. Zhou Kangqu et al solve the problem of strong noise in the bearing vibration signal by combining wavelet threshold with EMD and obtain good effect. Although the above noise reduction methods have certain effects, they have certain drawbacks, such as easy occurrence of modal aliasing and end-point effects in the decomposition process. Gu Ya et al decompose the collected rolling bearing vibration signals by using EEMD, so that noise existing in the rolling bearing vibration signals is effectively removed. However, EEMD has a problem that it cannot completely remove the interference signal in the vibration signal with respect to EMD, and also causes an increase in calculation time. Aiming at the problem of noise reduction, the L.Donoho firstly proposes a wavelet transformation method, and adopts soft and hard threshold values to filter and reduce noise for wavelet coefficients in the decomposition process. Pengfei Liang et al propose a new method WT-IResNet for signal noise reduction. In the current signal noise reduction method, wavelet transformation has good effect, but in the wavelet decomposition process, threshold setting is a key step, and is usually selected by an experimental or manual experience mode, and the self-adaption capability of a model is poor because the wavelet transformation cannot be adjusted according to different signals. The VMD solves the problems of modal aliasing, end-point effect and the like to a certain extent, can effectively separate components and realize self-adaptive frequency domain separation of signals, and mainly obtains the bandwidth and the frequency center of each IMF. The VMD method is used by mu et al to reduce noise, which is a problem of large interference to the external environment of the bearing and large noise component. Wu Luming et al propose a VMD-based Independent Component (ICA) algorithm based on the fact that the fault signal generated in the gearbox is very weak and very susceptible to interference from factors such as external environmental noise. However, VMD has a certain problem that two critical parameter modal decomposition numbers and secondary penalty factors need to be set manually before decomposition, and the subjective performance and the universality are poor. The methods of how to reasonably select these two parameters have been studied by the scholars. Bian Jie uses genetic algorithms to adaptively optimize the number of model decompositions and the quadratic penalty factors. Zhang Jun et al use a particle swarm optimization algorithm to optimize both parameters. However, these two algorithms also have problems of slow solving speed and low precision. The whale optimization algorithm (whale optimization algorithm, WOA) is a group intelligent optimization algorithm which is inspired by Seyedali Mirjalili teaching and receiving whale social behaviors in 2016, and has the characteristics of simple inspiring mechanism, few control parameters, high convergence rate, strong global searching capability and the like when solving some complex optimization problems.

The rolling bearing state detection and fault diagnosis technology is new along with the progress of the age from the beginning of the occurrence, and the rolling bearing fault diagnosis technology also enters the peak period of development. The machine learning research is continuously rising, and the deep learning theory is gradually becoming a hot learning algorithm in the field of machine learning. Guo et al propose a diagnostic method for directly classifying a continuous wavelet transform scale map of rolling bearing vibration signals by using a convolutional neural network. Tamilselvan et al propose a fault diagnosis method based on a Deep Belief Network (DBN) for performing fault diagnosis on an aircraft engine and a power transformer respectively. Yuan Jianhu et al extract the time-frequency diagram of the rolling bearing signal as input to the two-dimensional CNN, enabling fault diagnosis under the CNN network architecture. However, these fault diagnosis methods have certain limitations, and do not mine the relation of data and the mutual dependence.

To solve this problem, it is proposed to display data in the form of an irregularity chart. Because the relationships between nodes are reflected on the edges of the connection in the graph data, the edge weights reflect the strength of the relationships, compared with the traditional method, the graph data needs to be established first, a certain complexity is improved, a serious challenge is brought to the standard neural network-based method, and some important operations (such as convolution) are easy to apply to the Euclidean domain, but the graph data is difficult to model in a non-Euclidean space.

The Graph Neural Network (GNN) is an artificial intelligence algorithm derived from graph theory, and can process graph data. Complex graph data is extended with new concepts and definitions under the influence of Convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), and the like, and generates corresponding graph convolutional neural networks (GCNs), graph Recurrent Neural Networks (GRNNs), and Graph Automatic Encoders (GAEs).

Since GNNs can model the interdependence between data and embed it into extracted features, this approach is becoming a research hotspot in the field of fault diagnosis. Gao et al, using a rotating machine as a study object, performing fault diagnosis on the device in combination with a graph convolution neural network, constructing local geometric properties among vibration samples, constructing all vibration samples into an undirected weighted k-nearest neighbor graph, performing fault diagnosis through a depth map neural network, and verifying the effectiveness of the method by using gear and bearing data. Zhang et al converts the acquired acoustic signals into a graph, and models the graph by using GCN, thereby realizing the fault diagnosis of the roller bearing. Yu et al first construct a graph dataset and then implement fault classification through the proposed fast depth GCN. Shao Haidong et al constructed a semi-supervised learning model based on node level graph attention network, further mining representative bearing failure characteristics through the attention mechanism. The validity of the method is proved by analyzing bearing fault experimental data under two groups of time-varying rotating speeds. Yang Chaoying et al designed a graph attention short-short time memory network for mining timing vibration signal characteristics and time dependency relationships implicit in graph characteristics of road graphs, thereby deeply reflecting the bearing life-cycle degradation process. These methods also have a problem in that the importance of the input information is ignored to some extent.

Disclosure of Invention

In order to improve the noise reduction effect and prevent the problems described in the background art, the invention provides a rolling bearing fault diagnosis method based on WOA-VMD and GAT. Aiming at the problem that an end effect and modal aliasing can occur in signal decomposition, the method uses the VMD to separate signals by using a fixed bandwidth, so that the problems of modal aliasing and end effect are solved to a certain extent, two key parameters in the VMD are optimized and determined by a WOA optimization algorithm, so that the model can adaptively decompose the signals, finally, a network structure model for GAT fault diagnosis is constructed, meanwhile, the node classification of graph structure data is executed by using an Attention (Attention) system structure, the weight distribution of sensitive information is improved, reasonable network structure parameters are selected, and the accuracy and precision of fault diagnosis are improved.

The invention aims at realizing the following technical scheme:

a rolling bearing fault diagnosis method based on WOA-VMD and GAT comprises the following steps:

step one, WOA-VMD-based signal decomposition and reconstruction

Carrying out self-adaptive determination on the number k of parameter modes of VMD decomposition and penalty parameter alpha through a WOA optimization algorithm, so as to carry out VMD decomposition on an original signal, and then screening IMF components with large correlation in the decomposed signal by adopting Pearson correlation analysis so as to reconstruct the signal and finish signal noise reduction;

step two, rolling bearing fault diagnosis based on graph attention neural network

And combining the Attention and graph convolution operation to construct a graph Attention neural network rolling bearing fault diagnosis model, and distributing more specific gravity to the value information for optimizing the information collection stage of the constructed graph and improving the fault diagnosis accuracy of the model.

Compared with the prior art, the invention has the following advantages:

1. according to the invention, the number k of VMD decomposition parameter modes and penalty parameter alpha are adaptively determined through a WOA optimization algorithm, so that VMD decomposition is carried out on an original signal, and then the IMF component with large correlation is screened out by adopting Pearson correlation analysis on the decomposed signal, so that the signal is reconstructed, the phenomenon of modal aliasing in the traditional modal decomposition is effectively solved, and the noise reduction effect is excellent.

2. According to the invention, the Attention and graph convolution operation is combined to construct a graph Attention neural network rolling bearing fault diagnosis model, and more specific gravity is allocated to the value information so as to optimize the information collection stage of the constructed graph. The graph attention network is an information gathering phase that applies attention mechanisms to optimize the build graph. The node classification of the graph structure data of the present invention employs an attention-based architecture that computes hidden representations of neighboring nodes of a node in the graph using self-attention mechanisms. Firstly, the original characteristics of the nodes and the neighborhood characteristics of the nodes are input into a softmax layer to obtain the attention coefficient of each node. Then, a new node characteristic node is obtained through the inner product of the generated attention coefficient and the corresponding adjacent node characteristic. Experimental verification shows that compared with a MLP, attention model and a GCN model, the method has the advantages of higher convergence rate, higher diagnosis precision and lower loss value.

Drawings

FIG. 1 is a flow chart of a WOA-VMD and GAT based rolling bearing fault diagnosis method;

FIG. 2 is a time-frequency domain diagram of a simulated signal;

FIG. 3 is a waveform diagram after noise addition;

FIG. 4 is a WOA-VMD iteration curve;

FIG. 5 is a time-frequency domain diagram of the IMF after WOA-VMD decomposition;

FIG. 6 is a time-frequency domain diagram of different noise reduction methods;

FIG. 7 is a Kaiser Chu Da bearing experimental platform;

FIG. 8 is a graph showing experimental versus diagnostic accuracy variation;

FIG. 9 is a graph showing experimental comparative loss value variation;

FIG. 10 is a confusion matrix for different fault diagnosis methods;

FIG. 11 is a schematic diagram of a BPS-bearing accelerated life and fault simulation experiment table;

FIG. 12 is a time-frequency domain plot of a vibration signal;

FIG. 13 is a time-frequency domain diagram of the IMF after WOA-VMD decomposition;

FIG. 14 is a WOA-VMD iteration curve;

FIG. 15 is a time-frequency domain plot of the vibration signal after noise reduction;

FIG. 16 is a graph showing experimental versus diagnostic accuracy variation;

FIG. 17 is a graph showing the variation of experimental comparative loss values.

Detailed Description

The following description of the present invention is provided with reference to the accompanying drawings, but is not limited to the following description, and any modifications or equivalent substitutions of the present invention should be included in the scope of the present invention without departing from the spirit and scope of the present invention.

The invention provides a rolling bearing fault diagnosis method based on WOA-VMD and GAT, which is characterized in that the WOA is used for adaptively determining VMD decomposition parameter punishment factors and decomposition layer numbers, vibrating signals are decomposed by VMD, signals with few noise components are screened out from a plurality of IMF components to reconstruct, and a rolling bearing fault diagnosis method based on a graph attention neural network (GAT) model is provided for mining the relation and dependence among data and improving the precision of fault diagnosis. A network structure model of GAT fault diagnosis is constructed, and meanwhile, the node classification of the graph structure data is performed by using an Attention (Attention) architecture, so that the weight distribution of sensitive information is improved. As shown in fig. 1, the method specifically includes the following steps:

step one, WOA-VMD-based signal decomposition and reconstruction

And carrying out self-adaptive determination on the number k of the parameter modes of VMD decomposition and the punishment parameter alpha through a WOA optimization algorithm, so as to carry out VMD decomposition on the original signal, and then screening IMF components with large correlation in the decomposed signal by adopting Pearson correlation analysis so as to reconstruct the signal and finish signal noise reduction.

The method comprises the following specific steps:

step one, self-adaptive determination of VMD parameters and signal decomposition: and initializing parameters of WOA, calculating an envelope entropy value corresponding to each whale individual, recording the optimal value, outputting an optimal parameter combination, and decomposing the original signal by using the optimal k and alpha.

Step two, IMF component screening and signal reconstruction: and calculating the correlation coefficient of the IMF component and the original signal by adopting a Pearson correlation coefficient method, and reconstructing the IMF component with the phase relation number exceeding a threshold value (0.6) to finish signal noise reduction.

Step two, rolling bearing fault diagnosis based on graph attention neural network

And combining the Attention and graph convolution operation to construct a graph Attention neural network rolling bearing fault diagnosis model, and distributing more specific gravity to the value information for optimizing the information collection stage of the constructed graph and improving the fault diagnosis accuracy of the model. The method comprises the following specific steps:

step two,: and importing the data after noise reduction, converting the vibration signal data into graph structure data through a K-nearest neighbor (KNN), normalizing, and dividing the data into a test set, a training set and a verification set.

Step two: the fault diagnosis model of the GAT rolling bearing is built by adopting a multi-head attention mechanism, important neighborhoods are given higher weight, the sensitivity of the model to data containing fault diagrams is improved, and network model parameters such as learning rate, batch size, iteration number and the like are set, wherein the method comprises the following steps:

the GAT rolling bearing fault diagnosis model is built by initializing the model and then setting model parameters in detail, wherein the number of convolution kernel layers is two, and the specific parameters are as follows: a first layer convolution kernel: [2048,2048], second layer convolution kernel: [2048,2048], first layer activation function: relu, second layer activation function: relu, first full connection layer: [2048,1024], second full tie layer: [2048,1024], node deactivation rate: 0.2, batch normalization: 1024, loss function: cross entropy, optimizer: random gradient descent, iteration number: 100, batch size: 64, learning rate: 0.01.

the nodes are assigned with weights, important neighborhoods are given higher weights, and the new node characteristics are obtained by the following processes: (1) Inputting node information in the constructed KNN undirected graph, and converting input features into high-order features by linear transformation; (2) Calculating the correlation degree of the node i and the node j in the adjacent domain to the target node; (3) Normalizing the soft max function in all j options to obtain a weight (attention) coefficient, and calculating by using the LeakyReLU as an activation function; (4) And calculating corresponding characteristic linear combinations, and obtaining final output characteristic vectors of each node through nonlinear activation function calculation after obtaining normalized weight (attention) coefficients. In addition, the invention mainly realizes the weight assignment through a multi-head attention mechanism (multi-head attention), and each head attention mechanism finally performs a summation and average processing on the output characteristic vectors, namely, the finally output characteristic vectors are spliced through mutually independent attention mechanisms.

Step two, three: and inputting the training set data into a GAT rolling bearing fault diagnosis model, training the model, obtaining output errors through a verification set, and updating network model parameters by back propagation of the errors.

Step two, four: and step two, repeating the step three, stopping updating the network model parameters after the termination condition is met, enabling the network model parameters to achieve the best effect, completing training, and then inputting a test set into a GAT rolling bearing fault diagnosis model to complete fault diagnosis of the rolling bearing.

Example 1: WOA-VMD-based signal noise reduction experimental verification

The present embodiment uses simulated fault signals to verify the method. The simulated vibration signal formula is as follows:

wherein: f (f) ₁ ＝80Hz，f ₂ ＝200Hz，f ₃ =300 Hz, the number of sampling points n=1024.

Simulation signal s ₁ 、s ₂ 、s ₃ The time and frequency domain diagrams are shown in fig. 2.

And then mixing the signals, adding Gaussian white noise n (t), wherein the Gaussian white noise is-10 dB, and the time domain diagram and the frequency domain diagram after noise addition are shown in figure 3. As can be seen from fig. 3, the mixed simulation signal is added to the noise n (t) to obtain a signal which is more realistic than the previous signal. It can be seen in the time domain that the range of variation increases and the difference in amplitude increases, whereas in the frequency domain, the amplitude in the figure is staggered, since the analog high-intensity noise is added. The noisy mixed signal is then input into the WOA-VMD model for decomposition to obtain a plurality of modal components and a frequency domain plot thereof. And optimizing the decomposition layer number and the penalty factor in the VMD algorithm through the WOA algorithm, and obtaining the optimal result.

As shown in fig. 4, the penalty factor α iteration curve, the penalty factor iteration enters a convergence state after the fifth time; an iteration curve of the optimal decomposition layer number k, and after the second iteration, searching an optimal solution; the iteration curve of the envelope entropy converges after the thirteenth iteration. Fig. 5 is a time-frequency spectrum diagram of a plurality of modal components decomposed from an optimal combination of parameters.

And then screening the modal components obtained by decomposition by using a Pearson correlation coefficient method, and reconstructing the correlation coefficient value between the modal components and the original signal with the correlation coefficient value larger than or equal to a set threshold value. And comparing EMD, EEMD, CEEMD, GA-VMD noise reduced signals, wherein the time-frequency domain diagram and the frequency domain diagram are respectively shown in figure 6. As can be seen from fig. 6, these methods have some effect on noise reduction, but the WOA-VMD has better noise reduction effect than the other two methods, and has obvious noise reduction effect on the whole frequency band.

Example 2: rolling bearing fault diagnosis comparison experiment verification based on graph attention neural network

The embodiment adopts bearing fault data in CWRU data set, and a Kassi Chu Da bearing experiment platform mainly comprises a three-phase asynchronous motor, a torsion sensor, an alternating current power dynamometer and a controller, as shown in fig. 7. The fault sizes are directly and sequentially set to be 0.1778mm, 0.3556mm and 0.5334mm, and the rotating speeds are 1797r/min, 1772r/min, 1750r/min and 1730r/min. The specific different fault conditions are shown in table 1, the driving end bearing is SKF6205, and the fan end bearing is SKF6203. And (3) carrying out electric spark machining on the single-point damage of the bearing, wherein the damage points of the outer ring of the bearing are arranged at three different positions of 3 points, 6 points and 12 points. And respectively placing acceleration sensors above the bearing seats of the motor end and the driving end, and collecting vibration acceleration signals of the fault bearing. Vibration signals are collected by a 16-channel data recorder, a bearing with the fault size of 0.1778mm and a bearing with the fault size of 0.3556mm are respectively positioned at the driving end and the fan end, the sampling frequency is set to be 12kHz, the bearing with the fault size of 0.5334mm is only placed at the driving end, and the sampling frequency is 48kHz.

The original data is first truncated by a window with length 1024 and taken as a small sample, and is regarded as a node, 10 nodes form a graph structure, 512 graph structures form a large sample, and 122 large samples are total. 60% of the total samples are randomly extracted as training sets, 20% as verification sets and 20% as test sets. The iteration number is 100, the initial learning rate is 0.01, and a random gradient descent (SGD) with momentum is adopted as an optimization algorithm, wherein the momentum of the SGD is 0.9. Batch scale 64, 100 cycles per model training for fault diagnosis, learning rate decay strategy was also used to adjust learning rate, and weight decay value was initialized to 0.0005; and an inactivation rate (dropout) is set, and the main function is to conceal a part of error weight and prevent overfitting.

TABLE 1 different fault state experimental parameter settings

In order to verify the superiority of the method provided by the invention, a plurality of fault diagnosis models are compared to show that the GAT model provided by the invention has very strong fault diagnosis capability.

From fig. 8 and fig. 9, it can be seen that the accuracy of the MLP, attention model reaches about 80% after 100 training iterations, but MLP, attention is not in a convergence state, so that the convergence speed of the graph neural network model is faster than that of the traditional neural network model, the diagnosis precision is higher, and the loss value of the graph neural network is obviously lower than that of the traditional neural network from the loss value change curve.

The accuracy of the GCN model and the GAT model after 100 training iterations is about 100%, the iteration time and the loss value of the GCN model are higher than those of the GAT model, and the diagnosis precision is not as good as that of the GAT model, so that the diagnosis result predicted by the GAT model method is more attached to the actual diagnosis result, the iteration speed is high, the model stability is good, the generalization capability is strong, and the problem of fault diagnosis of the rolling bearing can be well solved.

In order to intuitively evaluate the classification effect of the model, a confusion matrix is adopted for visual analysis. The diagnosis accuracy rates of the different fault diagnosis methods are shown in table 2 according to the classification condition of each fault class, and the confusion matrix of the test samples of the different fault diagnosis methods is plotted as shown in fig. 10.

Table 2 test set diagnostic accuracy for different algorithms

As can be seen from fig. 10 and table 2, the Attention method has high accuracy in diagnosing moderate rolling element faults and severe inner ring faults, but has poor effect in diagnosing faults when the accuracy in diagnosing mild rolling element faults and severe rolling element faults does not reach 50%; the MLP method has high diagnosis accuracy for moderate rolling body faults and severe inner ring faults, but has about 50% of diagnosis accuracy for mild rolling body faults and severe rolling body faults, and has poor diagnosis effect; the GCN method has good diagnosis effect on most faults, but has deviation on severe outer ring fault diagnosis, and the diagnosis effect is superior to that of MLP and Attention; GAT is very accurate for diagnosing various faults, and the diagnosis effect is superior to that of other three methods.

Example 3:

in the embodiment, a bearing fault diagnosis experiment table shown in fig. 11 is adopted to carry out fault diagnosis tests of various rolling bearings under different working conditions, so as to obtain fault diagnosis monitoring data of the test bearings. Wherein: the bearing model was ER-16K and the detailed parameters are shown in Table 3.

Table 3 parameters of the test bearings

The damage of the fault bearing is artificial damage, and the faults of the inner ring and the outer ring are that the laser marking machine is used for laser with the same groove length in the ball rolling grooves of the inner ring and the outer ring of the bearing; the rolling element faults are punched on the rolling element using a laser marking machine. The injury degree is divided into mild injury, moderate injury and severe injury, the width of the mild injury of the inner and outer circles is 0.1mm, the width of the moderate injury is 0.3mm, and the width of the severe injury is 0.5mm. The fault diagnosis experiment is provided with 3 different loads, and different fault positions, different damage degrees and different experiment rotating speeds are set under the three different loads. In addition, there are three sets of health bearing data, three under load.

The data needs to be subjected to noise reduction treatment before use, and the original data is subjected to signal decomposition and reconstruction by adopting a WOA-VMD signal decomposition and reconstruction method. Then randomly selecting a group of data, carrying out signal decomposition and reconstruction processing, and observing the feasibility of the method. A time domain plot and a frequency domain plot of the vibration signal are shown in fig. 12.

In the experiment, the severe fault of the inner ring with the rotating speed of 900r/min under the load of 100kg is selected as experimental data, the set sampling frequency is 25.6kHz, the noise exists in the data as can be seen from fig. 13, the WOA-VMD is used for noise reduction treatment, the obtained iteration curve and a plurality of IMF components obtained through decomposition are adopted, and the steps are shown in fig. 13 and 14.

And (3) carrying out correlation analysis on the IMF components decomposed by the WOA-VMD and the original signals, calculating correlation coefficient values of each IMF component one by one, screening out the IMF components for reconstruction, and completing the noise reduction process of the signals, wherein fig. 15 is a time-frequency domain diagram of the signals after the WOA-VMD is subjected to noise reduction. As can be seen from fig. 15, the noise reduction effect of the high-frequency part of the noise reduced signal is very obvious, a large amount of noise is removed, and the effective value information is reserved, so that the ideal noise removal effect is achieved, and preparation is made for subsequent fault diagnosis. The vibration signal is noise-reduced by using the WOA-VMD algorithm, and then the processed data is fault-diagnosed.

The signals processed by the WOA-VMD algorithm are used as input, the built network model and the structural parameters are used for dividing the input data according to the method, and a comparison test is carried out, as shown in figures 16 and 17. From fig. 16, it can be seen that the MLP model and the Attention model, which are input in the form of conventional data, have lower diagnostic accuracy than the GCN model and the GAT model, which are input in the form of graph data structures, and that the graph data structures used herein have higher diagnostic accuracy. The GAT model diagnostic accuracy is higher than the GCN model, and as can be seen from FIG. 17, the GAT model loss value is lower than the GCN model. Therefore, the GAT model has certain diagnosis stability and good accuracy.

The diagnostic accuracy of the different fault diagnosis methods according to the classification of each fault class is shown in table 4. It can be seen that the accuracy of the diagnosis of the GAT fault signals can reach 100%, the classification of the GCN to the medium outer ring faults is not accurate, the diagnosis of the MLP to the heavy rolling body faults and the medium inner ring faults is not accurate, and the diagnosis of the Attention to the heavy rolling body faults and the medium inner ring faults is not accurate. From the above, it can be seen that the diagnostic effect of the graph neural network model is better than that of other models, and the GAT model has better diagnostic stability than the GCN model. Therefore, the superiority of the GAT used in the invention is proved by the indexes of accuracy and precision.

Table 4 test set diagnostic accuracy for different algorithms

Claims (4)

1. A rolling bearing fault diagnosis method based on WOA-VMD and GAT is characterized by comprising the following steps:

step one, WOA-VMD-based signal decomposition and reconstruction

Carrying out self-adaptive determination on the number k of parameter modes of VMD decomposition and penalty parameter alpha through a WOA optimization algorithm, so as to carry out VMD decomposition on an original signal, and then screening IMF components with large correlation in the decomposed signal by adopting Pearson correlation analysis so as to reconstruct the signal and finish signal noise reduction;

step two, rolling bearing fault diagnosis based on graph attention neural network

And combining the Attention and graph convolution operation to construct a graph Attention neural network rolling bearing fault diagnosis model, and distributing more specific gravity to the value information for optimizing the information collection stage of the constructed graph and improving the fault diagnosis accuracy of the model.

2. The rolling bearing fault diagnosis method based on WOA-VMD and GAT according to claim 1, characterized in that the specific steps of the step one are as follows:

step one, self-adaptive determination of VMD parameters and signal decomposition: carrying out parameter initialization on WOA, calculating an envelope entropy value corresponding to each whale individual, recording the optimal value, outputting an optimal parameter combination, and decomposing an original signal by using optimal k and alpha;

step two, IMF component screening and signal reconstruction: and calculating the correlation coefficient between the IMF component and the original signal by adopting a Pearson correlation coefficient method, and reconstructing the IMF component with the phase relation number exceeding a threshold value to finish signal noise reduction.

3. The rolling bearing failure diagnosis method based on WOA-VMD and GAT according to claim 2, characterized in that the threshold value is 0.6.

4. The rolling bearing fault diagnosis method based on WOA-VMD and GAT according to claim 1, characterized in that the specific steps of the second step are as follows:

step two,: importing the data after noise reduction, converting the vibration signal data into graph structure data through a K-neighbor method, normalizing the graph structure data, and dividing the data into a test set, a training set and a verification set;

step two: a multi-head attention mechanism is adopted to build a GAT rolling bearing fault diagnosis model, important neighborhoods are given higher weight, the sensitivity of the model to data containing fault diagrams is improved, and network model parameters are set;

step two, three: inputting training set data into a GAT rolling bearing fault diagnosis model, training the model, obtaining output errors through a verification set, and updating network model parameters by back propagation of the errors;

step two, four: and step two, repeating the step three, stopping updating the network model parameters after the termination condition is met, enabling the network model parameters to achieve the best effect, completing training, and then inputting a test set into a GAT rolling bearing fault diagnosis model to complete fault diagnosis of the rolling bearing.