Rolling bearing fault diagnosis method based on deep learning

Abstract

The invention provides a rolling bearing fault diagnosis method based on deep learning, and relates to the technical field of mechanical fault diagnosis. According to the method, global optimization is carried out on a parameter combination in an ICEEMDAN decomposition process through a CPO optimization algorithm, the optimal amplitude of white noise and the number of times of adding the white noise are determined, an intrinsic mode function is screened and denoised in combination with a correlation coefficient, reconstruction of a vibration signal of a rolling bearing is completed, and the vibration signal of the rolling bearing is reconstructed. The problem of mode aliasing is solved while the reliability and the stability of a vibration signal of the rolling bearing are improved; a K nearest neighbor method is used to construct a graph model, an edge connection relation between nodes can be effectively established, so that a foundation is laid for subsequent deep analysis and processing of vibration signals based on a graph structure, a graph attention neural network is input, topological structure information of the nodes and edges is fully utilized, extracted node features are classified and identified, and the accuracy of vibration analysis is improved. Therefore, fault diagnosis of the rolling bearing is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a rolling bearing fault diagnosis method based on deep learning. Background Art

[0002] Rolling bearings are widely used in the mechanical field due to their advantages such as high speed, high efficiency and low noise. However, due to the harsh working environment, the probability of failure is high. At the same time, nearly 30% of the failures of relevant mechanical equipment are related to bearing damage. Since fault detection is based on physical quantities, and the vibration-based measurement method is low-cost, easier than direct observation, and more sensitive to external interference, it is often used in actual engineering, so the vibration-based fault detection method is adopted.

[0003] The presence of noise will affect the performance of mechanical equipment and the prediction results of equipment health, so how to reduce or avoid noise is particularly important. To address this problem, empirical mode decomposition (EMD) has been widely used in signal processing problems due to its good decomposition effect. Jia et al. used ensemble empirical mode decomposition (EEMD) to decompose the collected rolling bearing vibration signal in the article "Denoising method for vibration signal of hobbased on grey criterion and EEMD[J]. Chin. J. Sci. Instrum, 2019, 40(5): 187–194." to effectively remove noise. Compared with EMD, it did not completely eliminate the interference signal in the vibration signal and increased the calculation time. At the same time, Jiang Fukang et al. proposed a complete ensemble empirical mode decomposition (CEEMDAN) algorithm in the article "Rolling bearing fault diagnosis based on CEEMDAN and CNN-LSTM [J]. Electronic Measurement Technology, 2023, 46(05): 72-77" to better solve the modal aliasing phenomenon existing in EEMD. However, the disadvantage is that Gaussian white noise is directly added during the decomposition process, which may lead to excessive introduction of noise. The emergence of improved completely adaptive noise ensemble empirical mode decomposition (ICEEMDAN) improves the noise addition process by adaptively adjusting the noise level added to the signal to improve the decomposition quality. Since the decomposition effect of the ICEEMDAN method requires the artificial setting of two key parameters, however, this artificial setting is subjective and has poor generalization, the selection of parameters is very important for subsequent diagnosis.

[0004] In recent years, convolutional neural networks have made rapid progress in image recognition, natural language processing and other fields. Although convolutional neural networks have achieved great success in processing data in Euclidean spaces such as images and texts, they are difficult to process non-Euclidean data in real-world problems. Given the universal representation of graphs and the powerful functions of convolutional neural networks, graph neural networks (GNNs) have become a hot topic in the field of rolling bearing fault diagnosis in terms of how to model and express the interdependencies between data and integrate them into feature extraction. Zhang et al. in "Intelligent acoustic-based fault diagnosis of roller bearings using adeep graph convolutional network[J].Measurement, 2020, 156(6):579-585." converted sound signals into graphs and used graph convolutional neural networks (GCNs) to model the graphs for rolling bearing fault diagnosis. Yu et al. first constructed a graph dataset in "Fault Diagnosis of Wind Turbine Gearbox Using a Novel Method of Fast Deep Graph Convolutional Networks[J].IEEE Trans. Instrum. Meas, 2021, 70(4):1-14." and then used fast deep GCN to achieve fault classification. The above method does not consider the importance of input information and the correlation between data. As the neighborhood changes, the degree of correlation between fault information will also change. If the same weight is given, some key information will be weakened or lost, thereby reducing the accuracy and effectiveness of the diagnosis results. Summary of the invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a rolling bearing fault diagnosis method based on deep learning. The CPO optimization algorithm is used to determine the two key parameters in ICEMMDAN, which solves the problem of modal aliasing while improving the reliability and stability of the signal. In addition, the method uses an attention mechanism method to classify nodes in graph structure data, and assigns different attention weights to different neighborhoods, thereby identifying more important rolling bearing fault information.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] The present invention provides a rolling bearing fault diagnosis method based on deep learning, which specifically includes the following steps:

[0008] S1: collecting rolling bearing vibration signals, and dividing the collected rolling bearing vibration signals into a number of rolling bearing vibration signal sub-samples of equal length using an overlapping sampling method;

[0009] S2: Establish a signal denoising algorithm based on CPO-ICEEMDAN, decompose the rolling bearing vibration signal sub-samples, obtain several IMF components, screen out the key IMF components, and obtain the reconstructed sub-sample signal after denoising;

[0010] S2.1: Decomposing the rolling bearing vibration signal subsamples by using the improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm, and obtaining the IMF components of the rolling bearing vibration signal subsamples;

[0011] S2.2: Establish a signal denoising algorithm based on CPO-ICEEMDAN to filter out key IMF components and obtain the denoised reconstructed sub-sample signal;

[0012] The method of establishing a signal denoising algorithm based on CPO-ICEEMDAN to screen out key IMF components is as follows: the ICEEMDAN algorithm is optimized using the crown porcupine optimization algorithm CPO, multiple parameters in the ICEEMDAN algorithm are selected as the parameter combination to be optimized, the fitness index is set, and the parameter combination to be optimized is iteratively optimized using CPO to obtain the parameter combination with the best fitness index, and the parameter combination is applied to the ICEEMDAN algorithm to screen out the key IMF components and obtain the reconstructed sub-sample signal after denoising;

[0013] The signal denoising algorithm based on CPO-ICEEMDAN includes the following steps:

[0014] (1) Randomly generate the initial population of the CPO algorithm. The individuals in the population are multiple parameters in the ICEEMDAN algorithm as the parameter combination to be optimized. Set the optimization range of each parameter in the ICEEMDAN algorithm to ensure that the parameters change within a reasonable range.

[0015] (2) The subsample of the rolling bearing vibration signal is used as the initial signal as the input of the signal denoising algorithm based on CPO-ICEEMDAN, the minimum envelope entropy value is used as the fitness function of the current parameter combination, the initial signal is decomposed into several IMF components using the ICEEMDAN algorithm, and the fitness function value corresponding to each IMF component is calculated;

[0016] (3) Determine whether the envelope entropy value of each IMF component reaches the termination condition. If the termination condition is met, stop the iteration; otherwise, return to (2) to continue the optimization;

[0017] (4) After the iteration, save the ICEEMDAN parameter combination corresponding to the minimum fitness function value;

[0018] (5) Applying the parameter combination to the ICEEMDAN algorithm to obtain an optimized ICEEMDAN algorithm, inputting the initial signal into the optimized ICEEMDAN algorithm for decomposition to obtain the best IMF component, and screening the key IMF component based on the correlation coefficient criterion to obtain the reconstructed subsample signal after denoising;

[0019] S3: construct a K nearest neighbor graph for the reconstructed sub-sample signal after denoising, establish a fault diagnosis data set based on the K nearest neighbor graph, and divide the fault diagnosis data set into a training set, a validation set, and a test set in proportion;

[0020] The specific method for constructing a K nearest neighbor graph for the reconstructed sub-sample signal after denoising is:

[0021] For the reconstructed sub-sample signal after CPO-ICEEMDAN decomposition and screening, each data point in the reconstructed sub-sample signal is regarded as a node, and the Euclidean distance formula is used to calculate the distance metric between each node. For the nth node y in the N data points contained in the reconstructed sub-sample signal n , n∈[1,N], calculate the distance d(y n ,y j ), j∈[1,N], j≠n, y j To divide y n For the jth node, according to the preset K value, select the K nearest neighbor nodes for each node, and build a graph model based on the connection relationship between these nodes. The weight of the connection edge is determined by the distance rule;

[0022] S4: Establish a GAT fault diagnosis model, use the fault diagnosis data set to train the GAT fault diagnosis model, update and save the optimal parameters of the GAT fault diagnosis model by back propagation, and obtain the trained GAT fault diagnosis model after determining the optimal parameters of the model through training;

[0023] The GAT fault diagnosis model introduces the attention mechanism to optimize the information collection stage based on the graph convolutional neural network GCN. In the GAT fault diagnosis model, the self-attention mechanism module is used to calculate the node y n Each neighbor node y z For n Contribution of nz , as shown in the following formula:

[0024] e nz =θ(W(y n ||y z )) (1)

[0025] Among them, “||” represents the concatenation of vectors, θ is the self-attention coefficient, and W is the learnable weight matrix;

[0026] Each adjacent node y z Normalize the contribution of the self-attention mechanism module to obtain the attention weight a nz , as shown below:

[0027]

[0028] Among them, s∈N i For node y n Neighbor node set N i The sth neighbor node in;

[0029] Use the leakyRelu activation function to activate the new weights with a nonlinear function in the linear layer to obtain the improved attention weight a in the self-attention mechanism module n ' z , as shown in the following formula:

[0030]

[0031] According to the improved attention weight in the self-attention mechanism module, node y is completed n The feature summation and update of node y n The fault characteristic representation h n ′, as shown in the following formula:

[0032]

[0033] Among them, h n ′ is node y n The fault characteristics are expressed as follows: z is the neighbor node y z The initial feature representation of , σ is the activation function;

[0034] Set the node y n The fault characteristic representation h n ′ is used as the output of the GAT fault diagnosis model, and is accurately compared with the validation set to obtain the output error. The output error is used to update and save the optimal parameters of the GAT fault diagnosis model through back propagation to obtain a trained GAT model;

[0035] S5: Input the vibration data of the rolling bearing to be diagnosed into the trained GAT fault diagnosis model, output the diagnosed fault information, and complete the fault diagnosis of the rolling bearing.

[0036] The beneficial effects of the above technical solution are as follows: the present invention provides a rolling bearing fault diagnosis method based on deep learning, which globally optimizes the parameter combination in the ICEEMDAN decomposition process through the CPO optimization algorithm, determines the optimal amplitude of white noise and the number of times white noise is added, and screens and denoises the intrinsic mode function in combination with the correlation coefficient to complete the reconstruction of the rolling bearing vibration signal, and uses the K nearest neighbor method to build a graph model, inputs the graph attention neural network to classify and identify the extracted node features, and then realizes the rolling bearing fault diagnosis. ICEEMDAN improves the signal decomposition quality by adaptively adjusting the noise level added to the signal; uses the CPO optimization algorithm to determine its key parameter combination, while improving the reliability and stability of the rolling bearing vibration signal, it solves the problem of modal aliasing. In view of the characteristics of the bearing signal presenting as a one-dimensional sequence, the use of the K nearest neighbor method to construct a graph structure has significant advantages, and can effectively establish the edge connection relationship between nodes, thereby laying the foundation for the subsequent in-depth analysis and processing of vibration signals based on the graph structure. This method has extremely critical significance and value in the relevant fields of processing one-dimensional sequence signals. The GAT model uses the Attention structure to perform node classification on graph data, thereby improving the weight distribution of sensitive information, making full use of the topological structure information of nodes and edges, and helping to mine complex associations and patterns in fault signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A flow chart of a rolling bearing fault diagnosis method based on deep learning provided by an embodiment of the present invention;

[0038] Figure 2 A flow chart of a signal denoising method based on CPO-ICEEMDAN provided in an embodiment of the present invention;

[0039] Figure 3 A data sample waveform diagram of an original signal in an embodiment of the present invention;

[0040] Figure 4 Parameter optimization curve diagram of ICEEMDAN based on the CPO algorithm provided in an embodiment of the present invention, wherein (a) is a fitness curve when envelope entropy is selected as the fitness function, and (b) is a curve showing the change of hyperparameters with the number of iterations when white noise amplitude weight (Nstd) and noise addition times (NE) are selected as the parameter combination to be optimized;

[0041] Figure 5 The original signal decomposition diagram based on CPO-ICEEMDAN provided in the embodiment of the present invention, wherein (a) is a time domain diagram of each decomposed signal, and (b) is a frequency domain diagram of each decomposed signal;

[0042] Figure 6The original signal based on CPO-ICEEMDAN and the time domain waveform, frequency domain waveform, and envelope spectrum of each component taking IMF1 and IMF3 components as examples provided in an embodiment of the present invention, wherein (a) is the time domain waveform, frequency domain waveform, and envelope spectrum of the original signal, (b) is the time domain waveform, frequency domain waveform, and envelope spectrum of the IMF1 component, and (c) is the time domain waveform, frequency domain waveform, and envelope spectrum of the IMF3 component;

[0043] Figure 7 A fault classification confusion matrix of the GAT fault diagnosis model provided by an embodiment of the present invention;

[0044] Figure 8 A comparison chart of the accuracy change curves of the GAT fault diagnosis model provided in an embodiment of the present invention, the CNN model, and the GCN model as a function of the number of iterations. DETAILED DESCRIPTION

[0045] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0046] This embodiment provides a rolling bearing fault diagnosis method based on deep learning, such as Figure 1 As shown, the specific steps include:

[0047] S1: collecting rolling bearing vibration signals, and dividing the collected rolling bearing vibration signals into a number of rolling bearing vibration signal sub-samples of equal length using an overlapping sampling method;

[0048] Overlapping sampling is a method of signal processing or data sampling. In the non-overlapping sampling method, the sampling window selects samples in sequence and there is no duplication between adjacent samples. The boundary of each sampling window may introduce some unnatural truncation effects. In the overlapping sampling method, the samples selected by adjacent sampling windows overlap to a certain extent during the sampling process, which helps to smooth the random fluctuations in the original signal. When the amount of original signal data is limited, overlapping sampling can effectively increase the number of samples, collect more abundant sample information, and better capture the local change characteristics of the signal. For high-frequency vibration signals with fast-changing characteristics, overlapping sampling can ensure that these local change details are not missed.

[0049] For a rolling bearing vibration signal sequence X=[X1,X2,…,X N ], set the sampling window size to w, the overlap ratio to p, select the first sampling window X1:w from the starting position of the collected rolling bearing vibration signal sequence, and sample the rolling bearing vibration signals X1 to X w; Calculate the starting position of the next sampling window according to the overlap ratio, and obtain the starting position of the next sampling window ζ=1+(1-p)w. Use the sampling window to sample the rolling bearing vibration signal in sequence until the end position of the sampling window exceeds the end of the signal, and obtain m rolling bearing vibration signal subsamples x=[x1,x2,…,x m ];

[0050] S2: Establish a signal denoising algorithm based on CPO-ICEEMDAN, decompose the rolling bearing vibration signal sub-samples, obtain several IMF components, screen out the key IMF components, and obtain the reconstructed sub-sample signal after denoising;

[0051] S2.1: Decomposing the rolling bearing vibration signal subsamples by using the improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm, and obtaining the IMF components of the rolling bearing vibration signal subsamples;

[0052] The improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm is a signal decomposition method based on empirical mode decomposition (EMD). By adding positive and negative complementary white noise to the signal to be decomposed, the residual noise in each component is reduced, the noise problem is effectively suppressed, and the shortcomings of easily generating false components and mode mixing are improved, thereby achieving a better denoising effect.

[0053] The ICEEMDAN algorithm decomposes the rolling bearing vibration signal subsample and white noise based on the EMD algorithm to obtain multiple IMF components of the rolling bearing vibration signal subsample and the IMF components of the white noise. By selecting the kth IMF component of the white noise as the noise component of the signal to be decomposed, the degree of noise introduction can be better controlled and the signal decomposition effect can be improved. Specifically, the following steps are included:

[0054] (1) Taking the subsample of the rolling bearing vibration signal as the initial signal, white noise is added to the initial signal to obtain the initial signal after adding noise:

[0055]

[0056] Among them, x i is the original signal; e1 is the noise standard deviation of the first decomposition signal; w (i) is white noise, E k (·) is the kth IMF component generated by EMD decomposition of the signal, E1(w (i) ) is the first IMF component obtained after white noise is decomposed by EMD;

[0057] (2) Calculate the first residual signal based on the local average of the initial signal after adding white noise

[0058]

[0059] Among them, <·> indicates the average value calculation;

[0060] (3) From the initial signal x i Subtract the first residual signal from the equation to get the first IMF component:

[0061] IMF1=x i -r1 (3)

[0062] (4) The second residual is estimated as r1+e2E2(w (i) ) and obtain the second IMF component:

[0063] IMF2=r1-r2=r1-<(r1+e2E2(w (i) ))> (4)

[0064] (5) Iterate step (4) to obtain the kth residual signal r k for:

[0065]

[0066] (6) According to the kth residual signal r k Get the kth IMF component IMF k :

[0067] IMF k =r k-1 -r k (6)

[0068] (7) When the preset number of IMF components is reached, the residual component meets the convergence criteria, or the number of iterations reaches the preset value, the signal after decomposition is terminated.

[0069]

[0070] S2.2: Establish a signal denoising algorithm based on CPO-ICEEMDAN to filter out key IMF components and obtain the denoised reconstructed sub-sample signal;

[0071] The method for establishing a signal denoising algorithm based on CPO-ICEEMDAN is as follows: the ICEEMDAN algorithm is optimized using the crown porcupine optimization algorithm CPO, multiple parameters in the ICEEMDAN algorithm are selected as the parameter combination to be optimized, the fitness index is set, the parameter combination to be optimized is iteratively optimized using CPO, and the parameter combination with the best fitness index is obtained. The parameter combination is applied to the ICEEMDAN algorithm, the key IMF components are screened out, and the reconstructed sub-sample signal after denoising is obtained;

[0072] Crested Porcupine Optimizer (CPO) is a meta-heuristic algorithm with fast convergence speed, short training time and low resource consumption, stable performance improvement and strong constraint processing ability. CPO algorithm is similar to other meta-heuristic population-based algorithms. It constructs a candidate solution set through population initialization and searches for the optimal solution from the candidate solution set consisting of C candidate solutions. The cth candidate solution Z in the candidate solution set is c for:

[0073]

[0074] Where N is the population size, L and U are the lower and upper limits of the search range, respectively, and r′ is a random number between 0 and 1;

[0075] CPO simulated various defense behaviors of crested porcupines (CP), which were divided into exploration and exploitation phases, including four defense strategies: visual, acoustic, odor, and physical attack. Among them, the first and second defense strategies (i.e., visual and acoustic) corresponded to the exploration phase of CPO, and the third and fourth defense strategies (i.e., odor and physical attack) corresponded to the exploitation phase of CPO;

[0076] 1) First defense strategy (visual):

[0077] When CP realizes the predator, it flaps its feathers. At this point, the predator has two choices: approach or move away. This behavior is modeled as:

[0078]

[0079] in, is the position of the jth CP in the tth iteration, is the best solution of the evaluation function for t, is a vector generated between the current CP and a CP randomly selected from the population, τ1 is a random number based on a normal distribution, τ2 is a random value in the interval [0,1], is the position of the predator at the tth iteration, as shown below:

[0080]

[0081] Where r is a random number between [1, N], used to represent a CP randomly selected from the population. is the position of the rth CP in the tth iteration;

[0082] 2) Second defense strategy (sound):

[0083] CP uses noise-making to threaten predators. When the predator approaches, the CP's voice becomes louder. This behavior can be modeled as:

[0084]

[0085] in, and are the positions of two randomly selected CPs in the tth iteration, τ3 is a random value in the interval [0,1], q is the position of the predator, which is between the current CP and a CP solution randomly selected from the population, and U1 is a binary vector that covers all possible probabilities of simulated sounds through randomly generated 0s and 1s;

[0086] 3) The third defense strategy (smell):

[0087] In the third defense strategy, CP secretes a foul odor that spreads in the area around it to prevent predators from approaching further. This behavior is modeled as:

[0088]

[0089] Among them, δ is a parameter used to control the search direction, γ t As a defense factor, are the odor diffusion factors, which are defined as:

[0090]

[0091] in, represents the objective function value of the j-th CP in the t-th iteration, ε is a small value to avoid division by zero, is a vector of randomly generated numbers between 0 and 1, rand is a variable that generates randomly generated numbers between 1 and 0, N is the population size, t is the number of iterations, and t max is the maximum number of iterations;

[0092] 4) The fourth defense strategy (physical attack):

[0093] When the predator is very close to the CP and attacks the CP, the CP will take physical attack, and the physical attack behavior is modeled as:

[0094]

[0095] in, represents the position of the jth CP at the tth iteration, represents the predator at that position, α is the convergence speed factor, τ4 is a random value in the interval [0,1], is the average force of the jth CP affecting the predator, and is defined as:

[0096]

[0097] Among them, m j is the mass of the jth predator, is the objective function, is the final velocity of the j-th predator at the next iteration t+1, and is allocated based on selecting random solutions from the current population, is the initial velocity of the j-th predator at iteration t, Δt is the number of the current iteration, and τ6 is a vector of random values ​​generated between 0 and 1.

[0098] CPO is used to optimize the parameter combination composed of multiple parameters in the ICEEMDAN algorithm. The white noise amplitude weight (Nstd) and the number of noise additions (NE) in the ICEEMDAN algorithm are selected as the parameter combination to be optimized. The envelope entropy is selected as the fitness function. Through continuous iterative calculation, the optimal parameter combination is obtained to further optimize the decomposition effect of the ICEEMDAN method and obtain the denoised reconstructed signal.

[0099] Envelope entropy is the entropy calculated by treating the amplitude distribution of the signal envelope as a probability distribution. Envelope entropy can reflect the complexity and uncertainty of the signal envelope. A lower envelope entropy indicates that the amplitude distribution of the envelope is relatively concentrated, and the signal may have a more regular modulation pattern; while a higher envelope entropy indicates that the amplitude distribution of the envelope is relatively dispersed, and the modulation pattern of the signal may be more complex and changeable. For example, for a simple sine wave modulation signal, its envelope is a constant and the envelope entropy is 0, because it has no uncertainty; while for a complex multi-frequency modulation signal, its envelope entropy will be higher. In the scenario where the signal takes the envelope entropy as the fitness function, it is usually in the optimization algorithm for signal processing (such as genetic algorithm, etc.), by adjusting the parameters of signal processing, the envelope entropy is optimized (such as maximum or minimum, depending on the specific problem), thereby optimizing the signal processing method.

[0100] The signal denoising algorithm based on CPO-ICEEMDAN is as follows Figure 2 As shown, the following steps are included:

[0101] (1) Randomly generate the initial population of the CPO algorithm. The individuals in the population are multiple parameters in the ICEEMDAN algorithm as the parameter combination to be optimized. Set the optimization range of each parameter in the ICEEMDAN algorithm to ensure that the parameters change within a reasonable range.

[0102] (2) The subsample of the rolling bearing vibration signal is input as the initial signal into the signal denoising model of CPO-ICEEMDAN, and the minimum envelope entropy value is used as the fitness function of the current parameter combination. The initial signal is decomposed into several IMF components using the ICEEMDAN algorithm, and the fitness function value corresponding to each IMF component, i.e., the envelope entropy value, is calculated;

[0103] (3) Determine whether the envelope entropy value of each IMF component reaches the termination condition. If the termination condition is met, stop the iteration; otherwise, return to (2) to continue the optimization;

[0104] (4) After the iteration, save the ICEEMDAN parameter combination corresponding to the minimum fitness function value;

[0105] (5) Applying the parameter combination to the ICEEMDAN algorithm to obtain an optimized ICEEMDAN algorithm, inputting the initial signal into the optimized ICEEMDAN algorithm for decomposition to obtain the best IMF component, and screening the key IMF component based on criteria such as correlation coefficient to obtain the reconstructed subsample signal after denoising;

[0106] S3 constructs a K nearest neighbor graph for the reconstructed sub-sample signal after denoising, establishes a fault diagnosis data set based on the K nearest neighbor graph, and divides the fault diagnosis data set into a training set, a validation set, and a test set in proportion;

[0107] The K-Nearest Neighbor (KNN) algorithm considers each fault data point in the reconstructed subsample signal as a node, finds the first K neighboring points for each node, and finds the nth node y among the N data points contained in the reconstructed subsample signal. n , n∈[1,N], the graph data is constructed based on the local neighbor information of the fault data point, which can accurately reflect the similarity and correlation between the fault data points. The calculation formula for the first l neighbor points of the fault data point is:

[0108]

[0109] in, Indicates that there are m samples, Ne(y n ) represents the fault data point y n Neighbor nodes of KNN generate a set Fault data point y n The Gaussian kernel weight function is used to calculate the fault data point yn The edge weights between neighboring nodes:

[0110]

[0111] Among them, e nz is the fault data point y n and neighbor node y z The edge weight between them, ξ is the bandwidth of the Gaussian kernel;

[0112] The specific method for constructing a K nearest neighbor graph for the reconstructed sub-sample signal after denoising is:

[0113] For the reconstructed sub-sample signal after CPO-ICEEMDAN decomposition and screening, each data point in the reconstructed sub-sample signal is regarded as a node, and the Euclidean distance formula is used to calculate the distance metric between each node. For the nth node y in the N data points contained in the reconstructed sub-sample signal n , n∈[1,N], calculate the distance d(y n ,y j ), j∈[1,N], j≠n, y j To divide y n For the jth node, according to the preset K value (K=5 in this embodiment), the K nearest neighbor nodes are selected for each node, and the graph model is constructed based on the connection relationship between these nodes. The weight of the connecting edge can be determined according to the distance rule, and the edge weight is set to the inverse of the distance to reflect the similarity or correlation between the nodes and reflect the local structural information between the data points;

[0114] S4: Establish a GAT fault diagnosis model, use the fault diagnosis data set to train the GAT fault diagnosis model, update and save the optimal parameters of the GAT fault diagnosis model by back propagation, and obtain the trained GAT fault diagnosis model after determining the optimal parameters of the model through training;

[0115] The GAT fault diagnosis model introduces an attention mechanism to optimize the information collection stage based on the graph convolutional neural network GCN. In the GAT fault diagnosis model, the self-attention module is used to calculate the node y n Each neighbor node y z For n The contribution of is as shown in the following formula:

[0116] e nz =θ(W(y n ||y z )) (twenty three)

[0117] Among them, “||” represents the concatenation of vectors, θ is the self-attention coefficient, and W is the learnable weight matrix;

[0118] Each adjacent node y z Normalize the contribution of the self-attention mechanism module to obtain the attention weight a nz , as shown below:

[0119]

[0120] Among them, s∈N i For node y n The neighbor node set N i The sth neighbor node in;

[0121] Use the leakyRelu activation function to activate the new weights with a nonlinear function in the linear layer to obtain the improved attention weight a in the self-attention mechanism module n ' z , as shown in the following formula:

[0122]

[0123] According to the improved attention weight in the self-attention mechanism module, node y is completed n The feature summation and update of node y n The fault characteristic representation h n ′, and the node y n The fault characteristic representation h n ′ is the output of the GAT fault diagnosis model, as shown in the following formula:

[0124]

[0125] Among them, h n ′ is node y n The fault characteristic representation is q z is the neighbor node y z The initial feature representation of , σ is the activation function;

[0126] Set the node y n The fault characteristic representation h n ′ is used as the output of the GAT fault diagnosis model, and is accurately compared with the validation set to obtain the output error. The output error is used to update and save the optimal parameters of the GAT fault diagnosis model through back propagation to obtain a trained GAT model;

[0127] S5: Input the vibration data of the rolling bearing to be diagnosed into the trained GAT fault diagnosis model, output the diagnosed fault information, and complete the fault diagnosis of the rolling bearing.

[0128] In this embodiment, the bearing fault data set from Case Western Reserve University (CWRU) in the United States with the drive end bearing model SKF6205 is selected as the fault diagnosis object. The experimental bench used for this data set includes an electric motor, a torque sensor / decoder, and a dynamometer. Using a sampling frequency of 12kHz, the vibration signals of the motor speeds are set to 1730r / min, 1750r / min, 1772r / min, and 1797r / min, respectively. The bearing is subjected to single-point damage using electrospark machining. The damage diameters are divided into 0.007 inches, 0.014 inches, 0.021 inches, 0.028 inches, and 0.04 inches (1 inch = 25.4mm). The damaged parts of the outer ring of the bearing are located at three o'clock, six o'clock, and twelve o'clock. The experimental parameter settings for different fault states of the bearing are shown in Table 1, where the sample for each state is 1024×400, with a total of 4000 sample data. The data sample waveform is shown in the figure below. Figure 3 shown.

[0129] Table 1 Experimental parameter settings for different fault states

[0130]

[0131] In this embodiment, the original vibration signal is divided into 400 segments with a length of 1024 sub-samples by overlapping sampling method. The parameter combination of ICEEMDAN is optimized by CPO optimization algorithm and combined with envelope entropy. The optimization results are shown in Figure 4 As shown, Figure 4 (a) is the fitness curve of ICEEMDAN. Figure 4 (b) is the curve of the change of ICEEMDAN parameter combination with the number of iterations; the optimal parameter combination is applied to the ICEEMDAN algorithm, and the rolling bearing vibration signal is used as the original signal to input into the optimized CPO-ICEEMDAN model. The ICEEMDAN decomposition result of the original signal containing noise is shown in Figure 5 As shown, Figure 5 (a) is the time domain diagram of each decomposed signal. Figure 5 (b) is the frequency domain diagram of each decomposed signal. As the IMF components increase, the original signal is divided into several components, and the corresponding frequencies decrease from top to bottom, indicating that the characteristic information contained in the IMF components is also decreasing. In the process of ICEEMDAN decomposition, there will be invalid components in the series of IMF components that are irrelevant to the characteristic information of the original signal. If the features of these components are directly extracted, not only will the workload increase greatly, but it will also have a certain impact on the analysis results.

[0132] In this embodiment, the Pearson correlation coefficient method is used to observe the degree of correlation between each component and the original signal, so as to accurately and efficiently complete the screening process. The Pearson correlation coefficient is a statistical indicator used to measure the degree of linear correlation between two variables, reflecting the strength and direction of the linear relationship between the variables. The Pearson correlation coefficient between each IMF component and the original signal is calculated. The Pearson correlation coefficient reflects the similarity between the IMF component and the original vibration signal. Selecting the IMF component with higher correlation to reconstruct the signal can retain its main characteristics. Set the correlation coefficient 0.1 as the threshold, the IMF above the threshold is retained, and the IMF below the threshold is discarded and does not participate in the reconstruction vector construction process. The specific values ​​of the correlation coefficients between each IMF component and the original signal are shown in the correlation coefficient table 2. It can be seen from Table 2 that with the increase of the IMF component, the correlation coefficient generally shows a downward trend, among which the correlation coefficients of IMF1, 2, 3, 4, and 5 are higher than 0.1, and the other IMF components are all less than this value. Therefore, IMF1~5 components are selected to complete the signal reconstruction process.

[0133] Table 2 Correlation coefficient table

[0134]

[0135] The original signal and the selected IMF1~5 components are converted into time domain, frequency domain and envelope spectrum respectively and analyzed. The time domain waveform, frequency domain waveform and envelope spectrum of the original signal and each component (taking IMF1 and IMF3 components as examples) are shown in Figure 2. Figure 6 As shown, Figure 6 (a) is the time domain waveform, frequency domain waveform and envelope spectrum of the original signal. Figure 6 (b) is the time domain waveform, frequency domain waveform and envelope spectrum of the IMF1 component. Figure 6 (c) is the time domain waveform, frequency domain waveform and envelope spectrum of the IMF3 component, Figure 6 (a) It can be seen that the signal waveform in the original signal is chaotic, which means that the signal contains too many interference components and it is difficult to extract the fault characteristic frequency; Figure 6 (b) and Figure 6 (c) It can be seen that the fault impact characteristics of the signal are obvious and more intuitive, indicating that the signal denoising model based on CPO-ICEEMDAN has effectively removed the interference components caused by environmental factors in the signal, and the reconstructed signal contains the main characteristic information of the fault.

[0136] The occurrence of modal aliasing is effectively solved by signal processing based on the CPO-ICEEMDAN signal denoising model. Subsequently, the IMF components with higher correlation in the decomposed signal are screened out through correlation analysis for signal reconstruction, and compared with other denoising methods by using evaluation indicators such as root mean square error and signal-to-noise ratio. The comparison results of different denoising methods are shown in Table 3. CPO-ICEEMDAN has the smallest root mean square error and the largest signal-to-noise ratio, indicating that the denoising effect is the best, verifying the superiority of this method.

[0137] Table 3 Comparison of denoising results using different denoising models

[0138]

[0139] In this embodiment, a GAT fault diagnosis model is established based on Pytorch, and the batch size, epoch, and learning rate are set to 64, 100, and 0.0001 respectively. The cross entropy loss function and Adam optimizer are selected as the hyperparameters of GAT. The model parameter settings are shown in Table 4. 10 in the output layer represents the number of fault data points in the first dimension K nearest neighbor graph and the feature dimension of the fault data points output by each layer in the second dimension. GAT convolution kernels 1 and 2 are used as feature extractors to extract the features of the fault data points, and a classifier is used to classify the fault features.

[0140] Table 4 GAT network structure and parameters

[0141]

[0142] After constructing the K nearest neighbor graph for the reconstructed vibration signal, it is input into the trained GAT fault diagnosis model to obtain the classification result and classification accuracy confusion matrix of the GAT fault vibration model as shown in Figure 7 As shown by Figure 7 It can be seen that the GAT fault diagnosis model correctly predicted labels 1, 3, 4, 5, 7, 8, and 10, and the classification accuracy on the test set reached 99.38%, indicating that the GAT fault diagnosis model can accurately extract the fault features of different rolling bearings and accurately classify them.

[0143] The comparison curves of different model accuracy and model iteration are as follows: Figure 8 As shown, the rolling bearing fault diagnosis method based on deep learning provided in this embodiment has a higher accuracy than the CNN model and the GCN model, and has a shorter convergence time. It has achieved a higher diagnostic accuracy in rolling bearing fault diagnosis and can effectively solve the problem of poor rolling bearing fault detection accuracy caused by modal aliasing.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on deep learning, characterized in that: The following steps are involved: S1: collecting rolling bearing vibration signals, and dividing the collected rolling bearing vibration signals into a number of rolling bearing vibration signal sub-samples of equal length using an overlapping sampling method; S2: Establish a signal denoising algorithm based on CPO-ICEEMDAN, decompose the rolling bearing vibration signal sub-samples, obtain several IMF components, screen out the key IMF components, and obtain the reconstructed sub-sample signal after denoising; S3: construct a K nearest neighbor graph for the reconstructed sub-sample signal after denoising, establish a fault diagnosis data set based on the K nearest neighbor graph, and divide the fault diagnosis data set into a training set, a validation set, and a test set in proportion; S4: Establish a GAT fault diagnosis model, use the fault diagnosis data set to train the GAT fault diagnosis model, update and save the optimal parameters of the GAT fault diagnosis model by back propagation, and obtain the trained GAT fault diagnosis model after determining the optimal parameters of the model through training; S5: Input the vibration data of the rolling bearing to be diagnosed into the trained GAT fault diagnosis model, output the diagnosed fault information, and complete the fault diagnosis of the rolling bearing.

2. The rolling bearing fault diagnosis method based on deep learning according to claim 1, characterized in that: Step 2 includes the following steps: S2.1: Decomposing the rolling bearing vibration signal subsamples by using the improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm, and obtaining the IMF components of the rolling bearing vibration signal subsamples; S2.2: Establish a signal denoising algorithm based on CPO-ICEEMDAN to filter out key IMF components and obtain the denoised reconstructed sub-sample signal.

3. The rolling bearing fault diagnosis method based on deep learning according to claim 2 is characterized in that: The signal denoising algorithm based on CPO-ICEEMDAN in step 2.2 includes the following steps: (1) Randomly generate the initial population of the CPO algorithm. The individuals in the population are multiple parameters in the ICEEMDAN algorithm as the parameter combination to be optimized. Set the optimization range of each parameter in the ICEEMDAN algorithm to ensure that the parameters change within a reasonable range. (2) The subsample of the rolling bearing vibration signal is used as the initial signal as the input of the signal denoising algorithm based on CPO-ICEEMDAN, the minimum envelope entropy value is used as the fitness function of the current parameter combination, the initial signal is decomposed into several IMF components using the ICEEMDAN algorithm, and the fitness function value corresponding to each IMF component is calculated; (3) Determine whether the envelope entropy value of each IMF component reaches the termination condition. If the termination condition is met, stop the iteration; otherwise, return to (2) to continue the optimization; (4) After the iteration, save the ICEEMDAN parameter combination corresponding to the minimum fitness function value; (5) The parameter combination is applied to the ICEEMDAN algorithm to obtain the optimized ICEEMDAN algorithm. The initial signal is input into the optimized CPO-ICEEMDAN algorithm for decomposition to obtain the optimal IMF component. Based on the correlation coefficient criterion, the key IMF components are screened to obtain the reconstructed subsample signal after denoising.

4. The rolling bearing fault diagnosis method based on deep learning according to claim 3 is characterized in that: The specific method of step 3 for constructing a K nearest neighbor graph for the reconstructed sub-sample signal after noise reduction is: For the reconstructed sub-sample signal after CPO-ICEEMDAN decomposition and screening, each data point in the reconstructed sub-sample signal is regarded as a node, and the Euclidean distance formula is used to calculate the distance metric between each node. For the nth node y in the N data points contained in the reconstructed sub-sample signal n , n∈[1,N], calculate the distance d(y n ,y j ), j∈[1,N], j≠n, y j To divide y n For the jth node, according to the preset K value, the K nearest neighbor nodes are selected for each node, and the graph model is constructed based on the connection relationship between these nodes. The weight of the connecting edge is determined by the distance rule.

5. The rolling bearing fault diagnosis method based on deep learning according to claim 4, characterized in that: The specific method of step 4 is: The GAT fault diagnosis model introduces the attention mechanism to optimize the information collection stage based on the graph convolutional neural network GCN. In the GAT fault diagnosis model, the self-attention mechanism module is used to calculate the node y n Each neighbor node y z For n Contribution of nz , as shown in the following formula: e nz =θ(W(y n ||y z )) (1) Among them, "||" represents the concatenation of vectors, θ is the self-attention coefficient, and W is the learnable weight matrix; Each adjacent node y z Normalize the contribution of the self-attention mechanism module to obtain the attention weight a nz , as shown below: Among them, s∈N i For node y n Neighbor node set N i The sth neighbor node in; Use the leakyRelu activation function to activate the new weights with a nonlinear function in the linear layer to obtain the improved attention weight a in the self-attention mechanism module n ' z , as shown in the following formula: According to the improved attention weight in the self-attention mechanism module, node y is completed n The feature summation and update of node y n The fault characteristic representation h n ′, as shown in the following formula: Among them, h n ′ is node y n The fault characteristics are expressed as follows: z is the neighbor node y z The initial feature representation of , σ is the activation function; Set the node y n The fault characteristic representation h n ′ is used as the output of the GAT fault diagnosis model and accurately compared with the validation set to obtain the output error. The output error is used to update the optimal parameters of the GAT fault diagnosis model through back propagation to obtain the trained GAT model.

Images