Bearing fault diagnosis method based on gadf-gan-avoa-cnn

Abstract

The application discloses a bearing fault diagnosis method based on GADF-GAN-AVOA-CNN, and steps are as follows: firstly, original one-dimensional bearing vibration signals collected are encoded into two-dimensional images by using GADF, and a part of the images are added with Gaussian white noise; then, a kind of multi-head self-attention mechanism is designed, and a GAN is used to denoise the two-dimensional fault images, so that the denoiser is used to replace the generator of the generative adversarial network, the noisy images are input into the denoiser to remove the noise, and the images output by the denoiser and the images without noise are input into the discriminator; in order to enhance the noise filtering effect, the multi-head attention mechanism is integrated into the denoiser, and when the denoiser and the discriminator are balanced in the game, the denoising effect reaches the best; finally, AVOA is used to optimize several key parameters in CNN, an optimal model structure and parameters are constructed, and the two-dimensional images after denoising are output to the optimized CNN for rolling bearing fault diagnosis. The application improves the accuracy and robustness of fault diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of bearing fault diagnosis, and particularly relates to a bearing fault diagnosis method based on GADF-GAN-AVOA-CNN. BACKGROUND

[0002] Rolling bearings are important parts in rotating machinery equipment, and their health directly affects whether the machinery equipment can operate normally. It is of great significance to keep rolling bearings safe and reliable in rotating machinery. According to statistics, the failure of rolling bearings accounts for 40% to 60% of the failure of rotating machinery equipment in actual industrial production. Therefore, it is particularly important to develop an efficient rolling bearing fault diagnosis method for industrial practice. Traditional bearing fault diagnosis methods mainly rely on expert databases and manual experience to extract fault features and determine fault types. Traditional bearing fault diagnosis methods mainly process original vibration signals and then extract features for analysis, including Hilbert-Huang transform, Fourier transform, wavelet transform, etc. These methods require experience to determine the specific fault type, which is not only time-consuming and labor-intensive but also has low accuracy. In the past few years, with the continuous development of deep learning, it has been widely applied in the field of bearing fault diagnosis. Convolutional neural networks (CNN), as a representative of deep learning, have made remarkable achievements in bearing fault diagnosis. They can learn representative features directly from raw data, avoiding the tedious process of extracting feature values and achieving end-to-end diagnosis.

[0003] In existing deep learning-based bearing fault diagnosis methods, the commonly used methods are one-dimensional signal processing. However, in variable speed conditions, one-dimensional signal feature extraction is difficult and there is a risk of feature loss. Meanwhile, considering that the most widely used and mature application of convolutional neural networks is two-dimensional images, the original one-dimensional signal is encoded into a two-dimensional image using GADF. Many deep learning fault diagnosis methods are designed for noise-free conditions, but the actual collected bearing vibration signals are often noisy. In a noisy environment, the accuracy of many deep learning fault diagnosis methods may decrease, so noise reduction is an essential step. In recent years, many scholars have attempted to use deep learning for image denoising, such as convolutional autoencoder, DnCNN, and deep residual shrinkage network. However, these denoising methods have problems such as poor denoising effect or low computational efficiency. Generative adversarial networks (GAN) are based on game theory and use the adversarial training between discriminators and generators to achieve denoising. When the discriminator cannot distinguish between the generated images and the labeled images, it means that the generated images have removed noise interference. To more effectively filter noise, GAN is combined with a multi-head self-attention mechanism.

[0004] CNN is an efficient means to solve the problem of bearing fault diagnosis, but it has many super parameters and complex structure parameters, and often needs to be repeated several times and "manual parameter adjustment" to determine the optimal model when building a CNN model. The "manual parameter adjustment" process is time-consuming and labor-intensive, and the reuse rate of the adjusted model parameters is not high, so it needs to be re-adjusted in different situations, so "manual parameter adjustment" limits the development of CNN in the field of fault diagnosis to some extent.

[0005] Therefore, the technical problem to be solved by the present application is to design a CNN model that can automatically determine the optimal super parameters and structure parameters, to overcome the defect that traditional deep learning cannot automatically extract the super parameters and structure parameters of the model, and to reduce the dependence on "manual parameter adjustment". In addition, it is necessary to further improve the accuracy and robustness of deep learning fault diagnosis under variable working conditions and noise environment. SUMMARY

[0006] In view of the technical problems existing in the background art, the bearing fault diagnosis method of GADF-GAN-AVOA-CNN provided by the present application strengthens the denoising effect of the image denoising model and overcomes the defect that traditional deep learning cannot automatically extract the super parameters and structure parameters of the model, and improves the accuracy and robustness of deep learning fault diagnosis under variable working conditions and noise environment.

[0007] In order to solve the above technical problems, the present application adopts the following technical solutions to realize:

[0008] A bearing fault diagnosis method based on GADF-GAN-AVOA-CNN, the steps are as follows:

[0009] Step 1: Use a vibration sensor to collect original one-dimensional vibration data of rolling bearings at different positions and different rotating speeds, and normalize the data to arrange a plurality of one-dimensional vibration time series signals containing three fault characteristics and normal state;

[0010] Step 2: Divide the arranged one-dimensional data into three parts according to the ratio of 7:2:1, and add 0-10dB Gaussian white noise to the first and third parts of data, the first part is the training set, the second part is the label, and the third part is the test set. Use the GADF (Gram Angle Difference Field) technology to convert all one-dimensional data into two-dimensional images that retain time correlation.

[0011] Step 3: Establish and initialize a generative adversarial network (GAN) based on a multi-head attention mechanism, which is used to denoise the noisy images in step 2: a typical generative adversarial network consists of two parts, namely the generator and the discriminator, and the denoiser is used instead of the generator in the generative adversarial network. In order to improve the feature extraction capability of the model and enhance the denoising effect, a multi-head self-attention module is added to the denoiser.

[0012] Step four: train the generative adversarial network based on the multi-head attention mechanism established in step three with the training set samples in step two, so that the implicit layer parameters are optimal, and the denoiser and discriminator reach a game balance, that is, the network's ability to remove noise is the strongest.

[0013] Step five: establish an African vulture optimization algorithm (AVOA) optimized convolutional neural network model (CNN): first initialize the CNN network structure; then determine the five key parameters of the CNN network, including the convolution kernel size m1 and the number of convolution kernels n1 of the convolution layer C1, the convolution kernel size m2 and the number of convolution kernels n2 of the convolution layer C3, and the learning rate c, train the CNN model to make its parameters optimal.

[0014] Step six: input the test set into the optimized GADF-GAN-AVOA-CNN model for rolling bearing fault diagnosis.

[0015] Further, the specific method of step one for collecting original vibration signals is: collecting original vibration data of the most common four states of rolling bearings, i.e., outer ring fault, inner ring fault, rolling element fault, and normal state. Normalize the data to obtain one-dimensional vibration time series signals of the four working states of the bearing.

[0016] Further, the specific method of step two for converting the training set and the test set into two-dimensional images is:

[0017] S1: first arrange the one-dimensional time series measured values X = {x1, x2, x3,..., xN} of the vibration signal in the [-1, 1] interval through formula (1) n

[0018]

[0019] S2: encode the data values as the cosine of the angle, encode the time as the radius, and re-express the converted data in S1 in polar coordinates through formula (2)

[0020]

[0021] In the formula: r i is the polar coordinate radius; t i is the sampling signal timestamp; N is a constant that normalizes the span of the polar coordinate system. As time moves forward, the values corresponding to the original signal are deformed between different angle points on the span circle in polar coordinates.

[0022] S3: finally, convert the one-dimensional vibration time series rearranged in S1 and S2 to the polar coordinate system, identify the correlation in different time intervals by calculating the angle difference between the corresponding points. The definition of GADF is as follows formula (3)​

[0023]

[0024] where I is a unit row vector; X 2 , X '2 are different row vectors, and GADF can preserve time correlation, so this method is selected for one-dimensional data encoding.

[0025] Further, the generative adversarial network based on the multi-head attention mechanism in step three is composed of two networks, a denoising network G and a discriminant network D, the noisy image in the training set in step two is taken as the input of the denoising network G, the denoiser outputs a preliminary denoised image x f and a label sample x r as the input of the discriminator; the discriminant network D discriminates whether the denoised data x f is true or false, through adversarial training, the denoiser G attempts to deceive the discriminator D with x f , while the discriminator D attempts to distinguish x r and x f , when the discriminator D identifies that the output image data x f is true, it indicates that the data x f output by the denoiser is extremely close to the label sample x r , that is, the noisy image data is denoised. In order to solve the training instability and mode collapse existing in the image denoising process of the generative adversarial network, and to improve the denoising ability, a multi-head self-attention module is added to the generative adversarial network.

[0026] Further, the multi-head self-attention module is added to the above generative adversarial network, and the specific improvement contents are as follows:

[0027] The multi-head self-attention mechanism is improved on the basis of the attention mechanism and can extract features from multiple dimensions, and its core expression is as follows

[0028]

[0029] The essence of the multi-head self-attention mechanism is the weighting method, here the original vibration information is given a high weight, and the noise information is given a low weight, in the application, a multi-head self-attention layer is added behind the last residual layer of the denoiser G of the generative adversarial network, so as to strengthen the denoising effect.

[0030] Further, in step five, the African vulture algorithm is used to adaptively extract 5 key parameters of the convolutional neural network CNN, and the specific steps are as follows:

[0031] S1: Determine the optimal vulture in the group: initialize the vulture population, and calculate the fitness of all vultures, compare and select the optimal solution as the first team's optimal vulture, the second optimal solution as the second team's optimal vulture, and other vultures approach the optimal vulture and the second optimal vulture through formula (5). Through iterative fitness, the entire vulture population is recalculated.

[0032]

[0033] wherein R i (t) represents the position of other vultures except the optimal vulture and the second optimal vulture; BestV1 represents the position of the optimal vulture; BestV2 represents the position of the second optimal vulture; L1 and L2 are parameters between [0, 1] and L1 + L2 = 1; p i represents the optimal vulture probability; f i represents the fitness of other vultures.

[0034] S2: Starvation rate of vultures: Vultures often search for food everywhere, if the vulture is in a satiated state, the energy is high, which can make them go further to forage; on the contrary, if hungry, the energy is low, the vulture does not have enough ability to fly further, and the hungry state will become aggressive. The mathematical modeling of this behavior of vultures is shown below.

[0035]

[0036] wherein F represents that the vulture is in a satiated state, T represents the maximum number of iterations, t represents the current iteration number, z is a random number between [-1, 1], h is a random number between [-2, 2]. rand1 represents a random value between [0, 1]. When z < 0, the vulture is hungry; if z ≥ 0, the vulture is full.

[0037] S3: Exploration phase: In the AVOA algorithm, vultures can search different random areas, a total of two different search strategies, one of which is selected by parameter P1. The parameter P1 must be assigned before the vulture search operation, and the assignment range is [0, 1], the following formula shows the use of the two strategies respectively.

[0038]

[0039] wherein P i (t) represents the position of the vulture at the tth iteration; P i (t+1) represents the position of the vulture at the t+1th iteration; X represents the random position of the vulture, X = 2 × rand, rand ∈ [0, 1], P1 represents the selection parameter; rand2 ∈ [0, 1], rand3 ∈ [0, 1], rand p1 ∈ [0, 1]; u bl b respectively represent the search region space upper and lower limits.

[0040] S4: development stage: the development stage has two strategies, namely the first development stage and the second development stage. By selecting parameters P2, P3, select which development strategy to update the vulture position. The first development stage position update formula is as follows:

[0041]

[0042] In the formula, rand4 [0, 1], rand5 [0, 1], rand6 [0, 1], rand p2 P2 represents the first development stage selection parameter, and other parameters have the same meaning as above.

[0043] The second development stage position update formula is as follows:

[0044]

[0045] In the formula, BestV1(t), BestV2(t) respectively represent the optimal position and the suboptimal position of the vulture at the tth iteration; rand p3 P3 represents the second development stage selection parameter; Levy(d) represents random walk, and its step length is heavy-tailed distribution, and the direction of the vulture is completely random and isotropic.

[0046] The patent can achieve the following beneficial effects:

[0047] 1、The original one-dimensional vibration signal is converted into a two-dimensional image by the Gramian angular difference field (GADF), compared with the one-dimensional signal, the two-dimensional image can fully play the advantages of the convolutional neural network in image feature extraction, and in the variable working condition, the complex factors caused by the change of the rotating speed do not need to be considered, and the data preprocessing is simple; compared with the fast Fourier transform, wavelet transform and other technologies, the GADF can retain the time correlation, and the high-dimensional features obtained by the convolutional neural network can make the time sequence approximate reconstruction.

[0048] 2、Compared with the existing image denoising technology, compared with the convolutional autoencoder, DnCNN, deep residual shrinkage network and the like, the generative adversarial network can avoid the loss of important information, retain the original information to the greatest extent, so as to improve the fault diagnosis accuracy; the generative adversarial network and the multi-head self-attention mechanism are combined, the noise data is given very low weight, and the original data is given very high weight, and finally the noise is effectively filtered out, and the method enhances the denoising ability.

[0049] 3、Compared with existing optimization algorithms such as particle swarm algorithm, grey wolf algorithm, genetic algorithm and the like, the African vulture optimization algorithm is a new meta-heuristic optimization algorithm, and has the advantages of strong optimization ability and fast convergence speed. Compared with simple'manual parameter adjustment', the use of the African vulture optimization algorithm to automatically determine the CNN network parameters improves the model diagnosis precision, overcomes the dependence of traditional CNN on artificial experience, avoids the problem of CNN falling into a local optimal solution and the like. BRIEF DESCRIPTION OF DRAWINGS

[0050] The application will be further described below in combination with the drawings and embodiments:

[0051] Figure 1 is a bearing fault diagnosis overall flowchart based on GADF-GAN-AVOA-CNN;

[0052] Figure 2 is a GADF one-dimensional data conversion two-dimensional image effect diagram;

[0053] Figure 3 is a generative adversarial network structure diagram based on the multi-head attention mechanism designed by the application;

[0054] Figure 4 is a generative adversarial network image denoising flowchart based on the multi-head attention mechanism designed by the application;

[0055] Figure 5 is a CNN self-adaptive parameter adjustment flowchart using the African vulture optimization algorithm provided by the application. DETAILED DESCRIPTION

[0056] Embodiment 1:

[0057] GADF refers to a gram angle difference field transformation; GAN refers to a generative adversarial network; AVOA refers to an African vulture optimization algorithm; CNN refers to a convolutional neural network

[0058] The preferred scheme is as shown in Figures 1 to 5 A bearing fault diagnosis method based on GADF-GAN-AVOA-CNN, the steps are:

[0059] Step 1: Sampling the original bearing vibration signal, the specific method is to collect original one-dimensional vibration data under different parts and different rotating speeds of the rolling bearing by using a vibration sensor, the bearing has four states of inner ring fault, outer ring fault, rolling element fault and normal working, and the data is normalized, and a plurality of one-dimensional vibration time sequence signals with different fault characteristics are arranged.

[0060] Step 2: The arranged one-dimensional data is divided into three parts according to the ratio of 7:2:1, and Gaussian white noise of 0-10 dB is added to the first and third parts of data. The first part is the training set, the second part is the label, and the third part is the test set. The Gram Angle Difference Field (GADF) technique is used to convert all one-dimensional data into two-dimensional images with time correlation. The specific steps of Gram Angle Difference Field (GADF) to encode the original one-dimensional time series into a two-dimensional image are as follows:

[0061] 2.1 First, normalize the one-dimensional time series vibration signal X = {x1, x2, x3,..., xN} to arrange it in the interval [-1, 1] by the following formula: n

[0062]

[0063] where x i represents the normalized vibration data; x i represents the original vibration data; max(X) represents the maximum value of the original vibration data; and min(X) represents the minimum value of the original vibration data.

[0064] 2.2 Each sampling point corresponds to a timestamp, and the data value is transformed into the cosine of the angle, and the timestamp is transformed into the radius. The time series signal normalized to [-1, 1] in 2.1 is transformed into polar coordinate form by the following formula:

[0065]

[0066] where r i is the polar coordinate radius; t i is the timestamp; N is a constant that normalizes the span of the polar coordinate system, represents the polar angle of the i-th timestamp polar coordinate. The normalized data is in the interval [-1, 1], and the corresponding cosine value is in the interval [0, π]. The cosine function is monotonic in [0, π], and for a given time series signal, there is only one polar coordinate corresponding to it, i.e., there is a unique inverse mapping.

[0067] 2.3 Finally, the one-dimensional time series vibration signal rearranged in steps 2.1 and 2.2 is transformed into the polar coordinate system, and the correlation between different time intervals is identified by calculating the angle difference between the corresponding points. The triangular difference between each point is calculated by the following formula:

[0068]

[0069] where I is the unit row vector; X 2 , X '2 are different row vectors, ​The angles corresponding to the i, j time points are represented respectively. The GADF can retain time correlation, and the converted image details are perfect without losing important information, the image color difference is obvious, and the features are easy to extract, so the GADF is selected to encode and convert the time sequence signal.

[0070] Step 3: Establishing a generative adversarial network based on a multi-head attention mechanism: The function of this network is to denoise the noisy image in step two: A typical generative adversarial network consists of two parts, namely a generator and a discriminator, and the denoiser is used instead of the generator in the generative adversarial network. In order to improve the feature extraction ability of the model and thus enhance the denoising effect, a multi-head self-attention module is added to the denoiser. The initial network structure is as follows:

[0071] 3.1 Denoiser: input layer, first convolutional layer + Relu activation layer, followed by 6 convolutional blocks, each consisting of a convolutional layer + batch normalization layer + Relu activation layer, a multi-head self-attention layer, and an output convolutional layer;

[0072] 3.2 Discriminator: input layer, first convolutional layer + batch normalization layer, then 5 groups of consecutive convolutional layers + batch normalization layers + Leaky-ReLU activation layers, a discrimination module, and an output layer.

[0073] Step 4: Training the initial generative adversarial network based on the multi-head attention mechanism established in step 3 using the training set samples in step 2 to optimize the hidden layer parameters, the specific steps are as follows:

[0074] 4.1 When training the generative adversarial network based on the multi-head attention mechanism, set the number of training samples each time as batch_size, set the maximum number of iteration steps as max_step, set the maximum number of iterations as max_epoch, set the denoiser learning rate as η, set the discriminator learning rate as γ, set the optimizer as Adam, and set the activation function as Relu. The denoiser completes m updates, and the discriminator completes 1 update.

[0075] 4.1 Combination Figure 3 , the first part of the training set is added with 0-10db Gaussian white noise to simulate noisy vibration samples in the actual environment, and the simulated noisy samples are input into the denoiser for preliminary denoising. The expression of the preliminary denoised sample is as follows:

[0076]

[0077] In the formula, x f denotes the denoised sample, y denotes the noisy vibration sample in the simulated actual environment, m denotes the noise distribution learned by the denoiser, res denotes the residual learning, φ denotes the model parameters, and G denotes the mapping between the noisy sample and the denoised sample in the denoiser.

[0078] 4.2 CombinationFigure 3 The preliminary denoising sample and the labeled sample are input into the discriminator, the "true or false" of the preliminary denoising sample is discriminated by the discriminator, if the discrimination is "false", the discrimination result is back propagated to the denoiser, the denoiser adjusts the hidden layer parameters, so that the denoiser effect is continuously optimized, at the same time the discrimination ability of the discriminator is also strengthened, the denoiser and the discriminator are finally satisfied with the iteration condition in the continuous confrontation training, so that the output result of the discriminator is "true", and the noisy sample is denoised.

[0079] 4.3 Multi-head self-attention mechanism

[0080] In combination Figure 4 It can be seen that the network structure of the generative adversarial network based on the multi-head attention mechanism and the embedding position of the multi-head self-attention mechanism, and the mathematical formula of the self-attention mechanism is:

[0081]

[0082] In the formula, I represents the collected bearing sample, W q ,W k ,W v respectively represent weight matrices; Q refers to query, K refers to key, and V refers to value;

[0083] The multi-head self-attention mechanism is composed of multiple self-attention modules, multiple sets of weight matrices Q, K and V are initialized, then the multiple sets of weight matrices are fused into one matrix, and finally multiplied by an initialized matrix, which is realized through the following formula.

[0084]

[0085] The learning position coding method is used to solve the limitation that the multi-head self-attention mechanism cannot capture position information, and specifically a nonlinear transformation is provided by FNN. The dimension of the multi-head self-attention mechanism output is determined by two parts. One is the product of the batch size and the sentence length of the input, and the other is the product of the number of convolution kernel layers and the size of the convolution kernel in the denoiser.

[0086] 4.4 Model loss function

[0087] The MSE of the second part of the training set sample (without noise) i.e. the labeled sample x r and the denoising sample x f is taken as the loss function of the denoiser, as shown below:

[0088]

[0089] In the formula, x i represents the labeled sample matrix, y i represents the denoising sample matrix, and m idenotes the noise distribution matrix learned by the denoiser.

[0090] The adversarial loss between the discriminator and the denoiser is a cross-entropy function, the principle of which is realized by the following formula:

[0091]

[0092] In the formula, x i denotes the label sample matrix, z denotes the noise matrix, G(z) denotes the denoiser network, denotes the expected value of the label sample, E z～p(z) denotes the expected value added with Gaussian white noise, and D(x) denotes the discriminator network.

[0093] In order to realize the optimization of model parameters, that is, to reduce the distance between the denoised sample and the label sample, the denoiser MSE loss function and the cross-entropy loss function of the denoiser and the discriminator are combined to construct a new model loss function, as follows:

[0094]

[0095] In the formula, a denotes a balance factor for balancing the denoiser MSE loss function and the cross-entropy loss function of the denoiser and the discriminator, and the other parameters have the same meaning as above.

[0096] 4.5 Evaluation index

[0097] The effect of filtering noise from the noisy image sample is good or not, that is, the similarity between the denoised image sample and the label sample is quantitatively evaluated by using PSNR (peak signal-to-noise ratio) and SSIM (structural similarity).

[0098]

[0099] In the formula, m and n respectively denote the length and width of the image, I(i,j) and K(i,j) respectively denote the pixel points between the two images, MAX I denotes the maximum value of the pixel point.

[0100] PSNR is evaluated according to the error between the corresponding pixel points of the two images, and the higher the PSNR value, the higher the image similarity, that is, the better the effect of filtering noise.

[0101]

[0102] In the formula, μ x and μ y are the average values of x and y, are the variances of x and y, σ xy is the covariance of x and y, and c1 and c2 are constants for maintaining stability.

[0103] SSIM measures the similarity of images from three different factors: brightness (mean), contrast (standard deviation) and structure (covariance). The value of SSIM ranges from -1 to 1, and the larger the value, the better the image denoising quality.

[0104] Step 5: Use the African vulture optimization algorithm to adaptively select the hyperparameters and structure parameters of CNN. The specific steps are as follows:

[0105] 5.1, combined Figure 5 , first determine the initial structure of CNN neural network, then optimize the hyperparameters of CNN to get the optimal CNN network, the hyperparameters of AVOA optimized CNN include the number and size of convolution kernel in convolution layer and learning rate.

[0106] 5.2, set the parameters of AVOA, initialize the vulture population, including the number of vulture population n, the maximum number of iterations T, the initial position of vulture population P i (i = 1, 2, 3,..., n), select parameters P1, P2, P3 for each stage update strategy, set the initial iteration t = 0, and establish the mapping relationship between the vulture population and the weights and thresholds of CNN network;

[0107] 5.3, calculate the fitness of all vulture populations and compare the results, evaluate the vulture population, find out the best fitness and the second best fitness as the optimal vulture position BestV1 and the second optimal vulture position BestV2, and update the remaining vulture positions.

[0108] 5.4, update the position of the remaining vultures using formula (5), calculate the hunger rate of the vultures using formula (6), if |F| ≥ 1, select formula (7) to update the position of the remaining vultures; if 0.5 ≤ |F| < 1, select formula (8) to update the position of the remaining vultures; if 0 ≤ |F| < 0.5, select formula (9) to update the position of the remaining vultures.

[0109] 5.5, after updating, calculate the fitness of all vulture populations and compare them, evaluate the vulture population, and take the best fitness and the second best fitness as the optimal vulture position BestV1 and the second optimal vulture position BestV2.

[0110] 5.5, let the current iteration number t = t + 1, if it meets the initial judgment condition, reaches a good enough position or the maximum number of iterations, then take the current African vulture population extreme value as the optimal solution, if it does not meet the set initial judgment condition, go back to step 5.5 and continue updating.

[0111] 5.6, the algorithm ends, output BestV1, get the optimal CNN hyperparameters and structure parameters.

[0112] Step 6: After the model training is completed, the test set samples are input into the optimized GADF-GAN-AVOA-CNN model for rolling bearing fault diagnosis.

[0113] The steps and advantages of the present application will be described in detail below in combination with embodiments:

[0114] Step 1: The original bearing vibration signal is sampled, the bearing model is 6205-2RS, the sampling frequency is 12Khz, the rotating speed is set to 1730r / m, the bearing is divided into inner ring fault, outer ring fault, rolling element fault and normal state four different working conditions, and the fault diameter of each working condition is divided into 0.007in, 0.014in and 0.021in (1in=2.54cm) three levels, so the bearing has a total of 10 states, including 9 fault states and 1 normal state. 1024 points are used as a sample by overlapping sampling, each different working condition state is divided into 500 samples, and a total of 5000 samples are collected.

[0115] Table 1 Rolling bearing fault types in CWRU data set

[0116]

[0117]

[0118] The arranged samples are divided into three parts according to 7:2:1, wherein the first part and the third part of data are added with 0-10dB Gaussian white noise, the first part is a training set, the second part is a label, and the third part is a test set. The original 5000 samples are converted into three kinds of two-dimensional images with different sizes 32x32, 64x64 and 128x128 by using GADF (Gramian Angular Difference Field), and the most suitable size is selected through test results.

[0119] Step 3: Establish a generative adversarial network based on a multi-head attention mechanism, and the network structure is as shown in Figure 4 The specific parameters are as follows:

[0120] Table 2 Denoiser parameter settings

[0121]

[0122] Table 3 Discriminator parameter settings

[0123]

[0124] Step4: Train the generative adversarial network based on the multi-head attention mechanism, set the number of training samples each time to 50, set the maximum iteration step to 10000, set the maximum iteration number to 20, set the denoiser learning rate to 0.01, set the discriminator learning rate to 0.001, set the training block size to 64x64, set the optimizer to Adam, and set the activation function to Relu. The denoiser completes 3 updates, the discriminator completes 1 update, the loss function balancing factor of the new model constructed in the present application is 0.08.

[0125] Step5: Use the African vulture optimization algorithm to adaptively select the hyperparameters and structure parameters of the CNN. First, determine the initial structure of the CNN. The embodiment selects the classic LeNet-5, which has the structure of convolutional layer C1, pooling layer S2, convolutional layer C3, pooling layer S4, convolutional layer C5, and fully connected layer F6.

[0126] The LeNet-5 parameters that need to be adaptively extracted by the AVOA are the number and size of the convolutional kernels in the convolutional layer C1, m1, n1, the number and size of the convolutional kernels in the convolutional layer C3, m3, n3, the number and size of the convolutional kernels in the convolutional layer C5, m5, n5, and the learning rate η, a total of 7 parameters.

[0127] The AVOA-optimized LeNet-5 parameters obtained by training the LeNet-5 model using the above step 6 are as follows

[0128] Table 4: AVOA-optimized CNN parameters

[0129]

[0130] Step6: After the model training is completed, under the same initial conditions, input the test set samples into the four different fault diagnosis methods, test each method 10 times, take the average of the diagnosis results, and finally obtain the fault diagnosis accuracy results as shown in Table 5. As can be seen from the bearing fault diagnosis results, the method of the present application has the highest fault diagnosis accuracy under noise conditions, proving that the method has higher accuracy and stronger robustness under noise conditions.

[0131] Table 5: Comparison of fault diagnosis accuracy of different methods

[0132]

[0133] The above embodiments are only preferred technical solutions of the present application, and should not be regarded as limiting the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, including equivalent replacement solutions of the technical features recited in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present application.

Claims

1. A bearing fault diagnosis method based on GADF-GAN-AVOA-CNN, characterized by: The steps include the following: Step one: collect original one-dimensional vibration data of different parts of rolling bearings under different rotating speeds using vibration sensors, and normalize the data to arrange a plurality of one-dimensional vibration time sequence signals containing three fault characteristics and a normal state; Step two: divide the arranged one-dimensional data into three parts according to a ratio of 7:2:1, wherein the first part and the third part of data are added with Gaussian white noise of 0-10 dB, the first part is a training set, the second part is a label, and the third part is a test set; convert all one-dimensional data into two-dimensional images that retain time correlation by using a Gram angle difference field technology; Step three: establish and initialize a generative adversarial network based on a multi-head attention mechanism, which is used to denoise the noisy images in step two: replace the generator in the generative adversarial network with a denoiser, and add a multi-head self-attention module to the denoiser to improve the feature extraction capability of the model and thus enhance the denoising effect, Step four: train the generative adversarial network based on the multi-head attention mechanism established in step three with the training set samples in step two, so that the implicit layer parameters are optimal, and the denoiser and the discriminator reach a game balance, that is, the network reaches the strongest ability to remove noise; Step five: establish a convolutional neural network model optimized by the African vulture algorithm: first initialize the CNN network structure; then determine five key parameters of the CNN network, including the convolution kernel size m1, the number of convolution kernels n1, the convolution kernel size m2, the number of convolution kernels n2, and the learning rate c, by using the African vulture optimization method, and train the CNN model to make the key parameters optimal; Step six: input the test set into the optimized GADF-GAN-AVOA-CNN model for rolling bearing fault diagnosis; The sub-steps of step five include: First, determine the initial structure of the CNN neural network, and then optimize the hyperparameters of the CNN by using the AVOA to obtain the optimal CNN network; Setting the parameters of AVOA, initializing the condor population, including the number of condor population n, the maximum number of iterations T, the initial position of the condor population P i (i = 1, 2, 3,..., n), updating the strategy selection parameters P1, P2, P3 at each stage, setting the initial iteration t = 0, and establishing the mapping relationship between the condor population and the weights and thresholds of the CNN network; Calculate the fitness of all vulture populations and compare the results, evaluate the vulture population, find the best fitness and the second best fitness as the optimal vulture position BestV1 and the second optimal vulture position BestV2, and update the remaining vulture positions; Update the positions of the remaining vultures, and after the update, calculate the fitness of all vulture populations and compare the results, evaluate the vulture population, and take the best fitness and the second best fitness as the optimal vulture position BestV1 and the second optimal vulture position BestV2; Let the current iteration number t = t + 1, if the initial judgment condition is met, the maximum number of generations is reached, the current African vulture population extreme value is taken as the optimal solution, if the initial judgment condition is not met, the update is continued; The algorithm ends, and outputs BestV1 to obtain the optimal CNN hyperparameters and structure parameters.

2. The GADF-GAN-AVOA-CNN-based bearing fault diagnosis method according to claim 1, characterized in that, In step one, the method for collecting original vibration signals is: collect original vibration data of the most common four states of rolling bearings, i.e., outer ring fault, inner ring fault, rolling element fault, and normal state; and normalize the data to obtain one-dimensional vibration time sequence signals of the four working states of the bearing.

3. The GADF-GAN-AVOA-CNN-based bearing fault diagnosis method according to claim 1, characterized in that, In step two, the method for converting the training set and the test set into two-dimensional images is: S1: first arrange the one-dimensional time series measured value X = {x1, x2, x3,..., xN} of the vibration signal in [-1, 1] interval through formula (1); n} through formula (1) to arrange it in [-1, 1] interval; x i represents normalized vibration data; x i represents raw vibration data; max(X) represents the maximum value of the raw vibration data; min(X) represents the minimum value of the raw vibration data; S2: The data value is encoded as the cosine of the angle, and the time is encoded as the radius. The converted data in S1 is represented in polar coordinates by formula (2): where: r i is the polar radius; t i is the sampling signal timestamp; N is a constant that acts to normalize the polar coordinate system span; as time progresses, the values corresponding to the original signal are transformed between different angular points on the span circle in polar coordinates; polar angle representing the polar coordinates of the i-th time stamp; S3: Finally, the one-dimensional vibration time series rearranged by S1 and S2 is converted into the polar coordinate system. By calculating the angle difference between the corresponding points, the correlation in different time intervals is identified. The definition of the Gramian angle difference field is as follows formula (3): where: I is a unit row vector; X 2 , X '2 are different row vectors, respectively; represents the polar angle of the jth time stamp polar coordinate.

4. The GADF-GAN-AVOA-CNN-based bearing fault diagnosis method according to claim 1, characterized in that, In step three, the generative adversarial network based on the multi-head attention mechanism is composed of two networks, a denoising network G and a discriminant network D. The noisy image in the training set of step two is taken as the input of the denoising network G, and the denoiser outputs a preliminary denoised image x f and the label sample x r is taken as the input of the discriminator; the discriminant network D discriminates the true and false of the denoised data x f . Through the adversarial training, the denoiser G attempts to deceive the discriminator D using x f , while the discriminator D attempts to distinguish x r from x f . When the discriminator D identifies the output image data x f as true, it indicates that the data x f output by the denoiser is extremely close to the label sample x r , that is, the denoising of the noisy image data is completed; in order to solve the training instability and the mode collapse existing in the image denoising process of the generative adversarial network and to improve the denoising ability, a multi-head self-attention module is added to the generative adversarial network.

5. The GADF-GAN-AVOA-CNN-based bearing fault diagnosis method according to claim 4, characterized in that, Add a multi-head self-attention module, the operation steps are as follows: The multi-head self-attention mechanism is improved on the basis of the attention mechanism, which can extract features from multiple dimensions. Its core expression is as follows The essence of multi-head self-attention mechanism is the weighting method, which gives high weight to the original vibration information and low weight to the noise information. Add a multi-head self-attention layer after the last residual layer of the denoiser G in the generative adversarial network, so as to strengthen the denoising effect; In the formula, W q ,W k ,W v respectively represent weight matrices; Q refers to query, K refers to key, and V refers to value.

6. The GADF-GAN-AVOA-CNN-based bearing fault diagnosis method according to claim 1, characterized in that, The steps of using African vulture algorithm to adaptively extract 5 key parameters of convolutional neural network CNN are as follows: S1: Determine the optimal vulture in the group: initialize the vulture population, calculate the fitness of all vultures, compare and select the optimal solution as the first team's optimal vulture, and the second team's optimal vulture as the second team's optimal vulture. Other vultures approach the optimal vulture and the second optimal vulture through formula (5); through iteration fitness, the whole vulture population is recalculated; where R i (t) denotes other vulture positions except the best and the second best vulture; BestV1 denotes the best vulture position; BestV2 denotes the second best vulture position; L1 and L2 are parameters between [0, 1] and L1 + L2 = 1; p i denotes the best vulture probability; f i denotes the fitness of other vultures; n denotes the number of vulture population; S2: Vulture hunger rate calculation: Vultures often look for food everywhere. If the vulture is in a state of satiety, the energy is high, which can make them go further to find food; on the contrary, if hungry, the energy is low, and the vulture does not have enough ability to fly further, and in the state of hunger, it will become aggressive; The mathematical modeling of this behavior of vultures is shown as follows: In the formula, F represents the vulture in a state of satiety, T represents the maximum number of iterations, t represents the current iteration number, z is a random number between [-1, 1], h is a random number between [-2, 2], rand1 represents a random value between [0, 1]; When z < 0, the vulture is hungry; if z ≥ 0, the vulture is full; S3: Exploration stage: In the AVOA algorithm, vultures can search different random areas, and there are two different search strategies, one of which is selected by parameter P1; The parameter P1 must be assigned before the vulture searches, and the value range is [0, 1]. The following formula shows the use of the two strategies respectively: where P i (t) denotes the position of the vulture at the tth iteration; P i (t+1) denotes the position of the vulture at the t+1th iteration; X denotes a random position of the vulture, X = 2 x rand, rand e [0, 1], P1 denotes a selection parameter; rand2 e [0, 1], rand3 e [0, 1], rand p1 e [0, 1]; u b , l b denote the upper and lower limits of the search region in space, respectively. S4: Development stage: The development stage has two strategies, namely the first development stage and the second development stage; Select the development strategy for vulture position update by selecting parameters P2 and P3; The position update formula of the first development stage is as follows: where rand4∈[0,1], rand5∈[0,1], rand6∈[0,1], rand p2 P2 represents a development first stage selection parameter; The position update formula of the second development stage is as follows: where BestV1(t), BestV2(t) represent the optimal and sub-optimal position of the vulture at the tth iteration respectively; rand p3 ∈ [0, 1]; P3 represents the selection parameter of the second phase of development; Levy(d) represents a random walk with step size of heavy-tailed distribution, the direction of the vulture is completely random and isotropic at each step.

Images