A Fault Diagnosis Method for Rolling Bearings

Abstract

The present invention provides a rolling bearing fault diagnosis method, which includes the following steps: dividing and converting the collected vibration signal of the rolling bearing into a wavelet time-frequency image; processing all wavelet time-frequency images to obtain a training set, a verification set and a test set; constructing a deep residual network; determining the number of wrongly labeled samples in the training set; using the training set and the verification set to train and verify the pre-created improved deep residual network to obtain a diagnostic model with the best diagnostic effect on the verification set; using the test set to test the optimal diagnostic model, and outputting the fault diagnosis result. The method uses singular value decomposition to solve the maximum value for dimensionality reduction, which can retain more eigenvalues; uses global singular values ​​to replace the fully connected layer to avoid overfitting; the loss function can maximize the reduction of the influence of the wrongly labeled samples on the diagnostic results, and the diagnostic model finally obtained has stronger adaptability to the noise environment, better fault diagnosis stability, and more accurate diagnostic results.

Description

Technical Field

[0001] The invention relates to the field of mechanical equipment state monitoring and fault technology, and in particular to a rolling bearing fault diagnosis method. Background Art

[0002] Rolling bearings are widely used in various modern industrial machinery such as automobiles, airplanes, fans, etc., and are essential key parts in many modern machinery. Due to the complexity and diversity of the external environment, causes of failure, failure modes and characteristics of rolling bearings, the existing rolling bearing fault diagnosis methods are still difficult to meet actual needs.

[0003] With the rapid development of big data technology, the application of deep learning technology represented by convolutional neural network in the field of rolling bearing fault diagnosis has received wide attention. This method first extracts features from the original vibration signal, and then uses the classification ability of convolutional neural network to perform fault diagnosis, and its fault diagnosis accuracy is high. Although many fault diagnosis methods based on deep learning have been studied and achieved good diagnostic results, the diagnostic performance of these methods usually depends on the quality of the signals collected under different health conditions. For signals obtained from actual industrial applications, their quality is usually difficult to guarantee. On the one hand, in actual industrial applications, there is inevitably a lot of noise in the collected signals. The convolution kernel and pooling kernel in the neural network usually act as local feature extractors. Under the interference of signal noise, the feature learning ability of the neural network is significantly affected because the local feature extractor has difficulty in detecting relevant fault features. On the other hand, in the field of fault diagnosis, adding data labels usually must be done by professionals, so it is costly to obtain high-quality labels, and the fuzzy boundaries of fault symptoms also make it more difficult to label fault data sets. Therefore, in real industrial applications, label errors caused by human factors are inevitable. Compared with the noisy environment (signal noise), noisy labels have a greater impact on the diagnostic performance of the classifier.

[0004] The main reasons are as follows:

[0005] 1) The importance of signal components varies. However, sample labels usually have the same impact on all samples.

[0006] 2) The size of signal features is much longer than that of labels. Therefore, in the presence of noisy labels or in noisy environments, the diagnostic performance of existing deep learning-based fault diagnosis methods will be significantly reduced, which limits their application in real scenarios.

[0007] Therefore, how to use deep learning technology to achieve high-precision fault diagnosis of rolling bearings in a noisy environment and with noise labels is an urgent problem to be solved. Summary of the invention

[0008] In order to solve the deficiencies of the above-mentioned prior art, the present invention provides a rolling bearing fault diagnosis method, which uses singular value decomposition to solve the maximum value instead of the dimensionality reduction of the pooling layer of the traditional diagnostic model, which can retain more eigenvalues ​​and make the fault diagnosis more accurate. And by solving the singular value of the fault signal image instead of the complex output vector of the fully connected layer of the traditional diagnostic model, overfitting is avoided and the diagnostic model is simplified to ensure the stability and accuracy of the diagnostic model. On the other hand, the method of the present invention proposes a new loss function, which can maximize the reduction of the influence of noise labels, i.e., wrongly labeled samples on the diagnostic results. It has strong adaptability to noise environments and noise labels, and can achieve high-precision fault diagnosis of rolling bearings in strong noise environments and when samples are incorrectly labeled.

[0009] Specifically, the present invention provides a rolling bearing fault diagnosis method, which comprises the following steps:

[0010] S1, dividing the collected vibration signal of the rolling bearing into a plurality of data segments connected one after another, and converting the collected vibration signal into a wavelet time-frequency image using a wavelet transform formula,

[0011] S2, randomly divide and normalize all wavelet time-frequency images to obtain training sets, validation sets and test sets;

[0012] S3. Determine the number of incorrectly labeled samples in the training set according to predetermined test indicators, where the test indicators include test accuracy and actual engineering experience;

[0013] S4. Construct a deep residual network, which includes a convolution layer, a singular value decomposition pooling layer and a singular value decomposition adaptive layer. The singular value decomposition pooling layer is used to obtain the singular values ​​of multiple convolution kernels using the singular value decomposition method, and construct a singular value matrix of multiple convolution kernels as the output matrix of the singular value decomposition pooling layer. The singular value decomposition adaptive layer is used to calculate the singular value of each wavelet time-frequency image using the singular value decomposition method, and concatenate the singular values ​​of all wavelet time-frequency images into a vector as the input vector of the deep residual network softmax layer;

[0014] S5. Add a loss function and an activation function to the deep residual network. The expression of the loss function is as follows:

[0015]

[0016] Among them, t1 and t2 are both constants, 0≤t1<1,t2>1, z j represents the input of the softmax layer, [·] +=max{·,0},γ t (z) by Calculate, N is the total number of training samples, x is the sample, i is the number of samples, c is the total number of fault categories, j is a constant of 1, 2, ... c;

[0017] The activation function is shown below:

[0018]

[0019] S6. Optimize the learning rate of the deep residual network and use a dynamic learning rate instead of the commonly used static learning rate so that the learning rate gradually decreases as the training progresses. The formula for the dynamic learning rate is as follows:

[0020]

[0021] Among them, μ is the current learning rate; μ min is the minimum learning rate; μ0 is the initial learning rate; δ is the decay factor; gs is the current training step number; ds is the ratio of the number of samples in the batch to the number of samples in the entire training set;

[0022] S7, using the training set to train the deep residual network, and using the verification set to verify the deep residual network of step S5, to obtain a trained fault diagnosis model with the best diagnostic effect verified by the verification set;

[0023] S8. Use the test set as the input of the optimal fault diagnosis model obtained in step S6, and output the final fault diagnosis result.

[0024] Preferably, the wavelet transform formula in step S1 is as follows:

[0025]

[0026] Among them, WT x (a,τ) is the wavelet transform coefficient, x(t) is the collected vibration signal, ψ a,τ (t) is the wavelet basis function, a is the scale factor, τ is the translation, specifically, the wavelet basis function is the Morlet wavelet, and the center frequency and bandwidth are both 3; t is the time for collecting the vibration signal.

[0027] Preferably, in step S1, a test platform is used to collect the vibration signal of the rolling bearing, and the test platform includes a motor, a torque sensor, a power meter and an electronic control device. The test platform includes a motor, a torque sensor, and a power meter that are respectively communicated with the electronic control device and controlled by the electronic control device; an acceleration sensor is used to collect the acceleration signal of the rolling bearing during operation, and the acceleration sensor is arranged along the axial or radial direction of the rotating axis of the rolling bearing.

[0028] Preferably, in step S1, Gaussian white noise is added to the collected vibration signal to simulate a strong noise environment, and the Gaussian white noise is divided into 8 levels according to the signal-to-noise ratio, which are -4, -2, 0, 2, 4, 6, 8, and 10 from low to high.

[0029] Preferably, in step S2, a plurality of wavelet time-frequency images are randomly divided into a training set, a validation set and a test set as samples according to a set division ratio, and normalization processing is performed on each of them.

[0030] Compared with the prior art, the effects of the present invention are as follows:

[0031] (1) In the process of constructing the fault diagnosis model, the present invention uses the maximum singular value instead of the traditional maximum pooling or average pooling to retain more eigenvalues ​​as much as possible, and the extracted feature map contains more feature information. Therefore, the final diagnosis model is more adaptable to the noisy environment, the fault diagnosis stability is better, and the diagnosis result is more accurate.

[0032] (2) The present invention sets a loss function containing two adjustable parameters in the diagnosis model to replace the traditional logistic loss function. The diagnosis model is more adaptable to noise labels. Even when some samples are incorrectly labeled, high-precision fault diagnosis of faulty bearings can still be achieved. This has very strong practical significance in industrial applications.

[0033] (3) The present invention adopts a new global singular value decomposition adaptive strategy to directly send the maximum singular value of each feature map into the softmax layer to replace the fully connected layer in the convolutional neural network for classification. This is not only more interpretable, but also can prevent the occurrence of overfitting and ensure the accuracy of the final result.

[0034] (4) The present invention optimizes the learning rate and uses a dynamic learning rate instead of the commonly used static learning rate. For the neural network model, the learning rate is a very important parameter that will directly affect the final diagnostic effect. The ideal learning rate should speed up the iterative update at the beginning of the training, and then slowly decay during the training process to prevent the network model from oscillating. Therefore, the text adopts a learning rate with a dynamic decay property, so that the learning rate gradually decreases as the training progresses, thereby ensuring the stability of the fault diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 is a flow chart of a rolling bearing fault diagnosis method according to an embodiment of the present invention;

[0037] Figure 3 is a schematic structural diagram of a rolling bearing in an embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of the improved deep residual network structure in an embodiment of the present invention;

[0039] Figure 5 It is a schematic diagram of singular value decomposition pooling in a rolling bearing fault diagnosis method according to an embodiment of the present invention;

[0040] Figure 6 It is a schematic diagram of the global singular value pooling principle in the rolling bearing fault diagnosis method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0042] The present invention provides a rolling bearing fault diagnosis method, such as Figure 1 As shown, it includes the following steps:

[0043] S1, dividing the collected vibration signal of the rolling bearing into a plurality of data segments connected one after another, and converting the collected vibration signal into a wavelet time-frequency image using a wavelet transform formula,

[0044] S2, randomly divide and normalize all wavelet time-frequency images to obtain training sets, validation sets and test sets;

[0045] S3. Determine the number of incorrectly labeled samples in the training set according to predetermined test indicators, where the test indicators include test accuracy and actual engineering experience;

[0046] S4. Construct a deep residual network, which includes a convolution layer, a singular value decomposition pooling layer, and a singular value decomposition adaptive layer. The singular value decomposition pooling layer is used to obtain the singular values ​​of multiple convolution kernels using the singular value decomposition method, and construct the singular value matrix of multiple convolution kernels as the output matrix of the singular value decomposition pooling layer. The singular value decomposition adaptive layer is used to calculate the singular value of each wavelet time-frequency image using the singular value decomposition method, and concatenate the singular values ​​of all wavelet time-frequency images into a vector as the input vector of the deep residual network softmax layer;

[0047] S5. Add loss function and activation function to the deep residual network. The expression of loss function is as follows:

[0048]

[0049] Among them, t1 and t2 are both constants, 0≤t1<1,t2>1, zj represents the input of the softmax layer, [·] + =max{·,0},γ t (z) by Calculate, N is the total number of training samples, x is the sample, i is the number of samples, c is the total number of fault categories, j is a constant of 1, 2, ... c;

[0050] The activation function is as follows:

[0051]

[0052] S6. Optimize the learning rate of the deep residual network and use a dynamic learning rate instead of the commonly used static learning rate so that the learning rate gradually decreases as the training progresses. The formula for the dynamic learning rate is as follows:

[0053]

[0054] Among them, μ is the current learning rate; μ min is the minimum learning rate; μ0 is the initial learning rate; δ is the decay factor; gs is the current training step number; ds is the ratio of the number of samples in the batch to the number of samples in the entire training set;

[0055] S7, using the training set to train the deep residual network, and using the verification set to verify the deep residual network of step S5, to obtain a trained fault diagnosis model with the best diagnostic effect verified by the verification set;

[0056] S8. Use the test set as the input of the optimal fault diagnosis model obtained in step S6, and output the final fault diagnosis result.

[0057] Preferably, the wavelet transform formula in step S1 is as follows:

[0058]

[0059] Among them, WT x (a,τ) is the wavelet transform coefficient, x(t) is the collected vibration signal, ψ a,τ (t) is the wavelet basis function, a is the scale factor, τ is the translation, specifically, the wavelet basis function is the Morlet wavelet, and the center frequency and bandwidth are both 3; t is the time for collecting the vibration signal. Specific embodiments

[0061] In order to facilitate the comparison of the rolling bearing fault diagnosis method of the embodiment of the present invention with other methods, the embodiments of the present invention all use the bearing experimental data of Case Western Reserve University in the United States. This data set is widely used in the field of fault diagnosis. However, it should be noted that the scope of application of the embodiment of the present invention is not limited to this data set.

[0062] This embodiment provides a rolling bearing fault diagnosis method, such as Figure 2 As shown, the following steps are included:

[0063] S1. Collect vibration signals of rolling bearings:

[0064] like Figure 3 The experimental platform used for data collection in this patent is shown, and the experimental platform mainly includes a 2-horsepower motor 1, a torque sensor 2, a power meter 3 and an electronic control device. The electronic control device communicates with other components and controls other components. The above components are all arranged on a base 4.

[0065] Single-point faults were arranged on the bearing using electrospark machining technology, with fault diameters of 0.007, 0.014 and 0.021 inches (1 foot = 2.54 cm), and the fault locations were on the inner ring, outer ring and rolling element, respectively, i.e. a total of 10 fault modes, namely: normal, rolling element 0.007, rolling element 0.014, rolling element 0.021, inner ring 0.007, inner ring 0.014, inner ring 0.021, outer ring 0.007, outer ring 0.014 and outer ring 0.021.

[0066] During the operation of the rolling bearing, an acceleration sensor can be used to collect acceleration signals, and the acceleration sensor is arranged along the axial or radial direction of the rotating shaft of the rolling bearing 5. For example, 12,000 sampling points are collected per second. In order to simulate a strong noise environment, Gaussian white noise is added to the collected vibration signal, wherein the Gaussian white noise is divided into 8 levels according to the signal-to-noise ratio, from low to high, respectively -4, -2, 0, 2, 4, 6, 8, 10.

[0067] S2. Divide the collected vibration signal into a plurality of data segments connected one after another according to the set segment length, and perform wavelet transformation on each data segment respectively to obtain a corresponding wavelet time-frequency image.

[0068] Specifically, the sampling points of each fault mode are divided into 200 data segments with a segment length of 600 sampling points, that is, each data segment contains 600 continuous sampling points, and the collected vibration signal is converted into a wavelet time-frequency image using the wavelet transform formula.

[0069] The collected vibration signal is converted into a wavelet time-frequency image by the following wavelet transform formula:

[0070]

[0071] Among them, WT x (a,τ) is the wavelet transform coefficient, x(t) is the collected vibration signal, ψ a,τ(t) is the wavelet basis function, a is the scale factor, and τ is the translation. Specifically, the wavelet basis function can be a Morlet wavelet, and the center frequency and bandwidth are both 3, and t is the time for collecting the vibration signal.

[0072] The vibration signal is converted into a wavelet time-frequency image through wavelet transform to reduce the influence of background noise, extract the fault characteristics of the vibration signal, and obtain the input sample format of the improved deep residual network.

[0073] According to the above method, the data of the sampling points of each segment of the 10 fault modes are converted to obtain the time-frequency image.

[0074] S3. After randomly dividing and normalizing all wavelet time-frequency images, we obtain the training set, validation set and test set.

[0075] Specifically, multiple wavelet time-frequency images are randomly divided into training set, validation set and test set as samples according to the set division ratio, and normalized respectively. For example, the wavelet time-frequency images of each fault mode are randomly divided into three parts: training set, validation set and test set with a division ratio of 7:2:1, and they are all normalized to [-1,1], that is, for each fault mode, the training set includes 140 wavelet time-frequency images (training sample number), the validation set includes 40 wavelet time-frequency images (validation sample number), and the test set includes 20 wavelet time-frequency images (test sample number), and each wavelet time-frequency image is a sample. The detailed data is shown in Table 1.

[0076] Table 1: Training set and sample set of wavelet time-frequency images for each fault mode

[0077]

[0078] S4. Determine the number of incorrectly labeled samples in the training set based on the predetermined test indicators, which include test accuracy and actual engineering experience. Test accuracy is defined as the ratio of all correctly diagnosed samples in the test sample to all test samples; actual engineering experience is based on the actual engineering situation. The incorrectly labeled sample data is caused by errors in the labeling work, so the number cannot be too large. Here we set the maximum ratio to 10% of the training samples.

[0079] The test indicators may also include parameters such as manual labeling cost. It is understandable that the fewer the number of incorrectly labeled samples, the higher the accuracy of the samples, and the higher the accuracy of the final fault diagnosis, but the higher the requirements for manual labeling, and the correspondingly higher the labeling cost.

[0080] According to Table 1, the total number of training set samples is 1400, so the number of mislabeled samples can be set to 0 (noise label ratio is 0), 28 (noise label ratio is 2%), 70 (noise label ratio is 5%) or 140 (noise label ratio is 10%).

[0081] S5. Use the training set to train the pre-created improved deep residual network, and select the best trained network model through the validation set. The improved deep residual network structure is shown in the figure below. Figure 4 shown.

[0082] The method for creating an improved deep residual network includes: designing a structure of the deep residual network based on a neural network model.

[0083] Specifically, the traditional deep residual network structure is improved, mainly including using the singular value decomposition dimensionality reduction method instead of the traditional maximum pooling or average pooling; using a new loss function containing two adjustable parameters instead of the traditional logistic loss function; using the global singular value decomposition adaptive strategy instead of the traditional fully connected layer.

[0084] like Figure 4 As shown, the deep residual network includes multiple convolutional layers, multiple singular value decomposition pooling layers and singular value decomposition adaptive layers. The singular value decomposition pooling layer is used to obtain the singular values ​​of multiple convolution kernels using the singular value decomposition method, and construct the singular value matrix of multiple convolution kernels as the output matrix of the singular value decomposition pooling layer. The singular value decomposition adaptive layer is used to calculate the singular value of each wavelet time-frequency image using the singular value decomposition method as the output vector of the deep residual network.

[0085] Figure 5 The figure shows a schematic diagram of the singular value decomposition pooling layer in this embodiment. The traditional maximum pooling or average pooling uses the maximum value or average value of the matrix for dimensionality reduction, which will lose a lot of feature information of the matrix. The singular value decomposition pooling layer uses the maximum singular value of the matrix for dimensionality reduction and outputs a matrix composed of the singular values ​​of multiple convolution kernels. On the basis of ensuring dimensionality reduction, it can not only retain more feature information, but also play a role in noise reduction.

[0086] Figure 6 The figure shows a schematic diagram of the global singular value adaptation strategy in this embodiment. The traditional fully connected layer contains many parameters, which easily leads to overfitting. The singular value decomposition adaptive layer of the present invention performs singular value decomposition on the entire feature map, and uses the maximum singular value as the eigenvalue of the neuron to convert a series of feature maps into a feature vector, thereby preventing overfitting and simplifying the model.

[0087] S6. In order to solve the problem of decreased fault diagnosis accuracy caused by noise labels, the present invention uses a loss function containing two adjustable parameters to replace the traditional logistic loss function. Compared with the traditional logistic loss function, the improved loss function has stronger robustness to outliers. The expression of the loss function is shown as follows:

[0088]

[0089] Among them, t1 and t2 are both constants, 0≤t1<1,t2>1, z j represents the input of the softmax layer, [·] + =max{·,0},γ t (z) by Calculation is performed, where N is the total number of training samples, x is the sample, i is the number of samples, c is the total number of fault categories, and j is a constant of 1, 2, ... c.

[0090] S7. Use the training set to train the deep residual network, and use the verification set to verify the deep residual network of step S5, so as to obtain a trained fault diagnosis model with the best diagnostic effect verified by the verification set.

[0091] S8. Input the samples in the test set into the trained IResNet model to obtain the fault diagnosis results.

[0092] In order to verify that the present invention has strong generalization, that is, the accuracy of the present invention in fault diagnosis has strong stability, we conducted 10 repeated experiments, and the average test accuracy of the 10 times is shown in Table 3.

[0093] Table 2 Average test accuracy of 10 trials

[0094]

[0095] In real-world environments, due to human error, wrong labels are inevitable. However, if there are too many noisy labels, the research will become meaningless, because such huge errors are not allowed in actual fault diagnosis. Therefore, in this paper, we added 0%, 2%, 5%, and 10% label noise rates to the training set to form four training sub-datasets. Each dataset class includes 1400 training samples, 400 validation samples, and 200 test samples. Our performance is improved by comparing the damage function of this patent with the original ResNet by using each dataset class. We repeated each experiment 10 times and used the average value of each experiment as the evaluation indicator, as shown in Table 2.

[0096] As shown in Table 2. As the LNR increases, the test accuracy of both methods decreases at different time periods. However, when the LNR is 0%, 2%, 5% and 10%, the performance of the model is better, which is 0.77%, 0.82%, 2.56% and 3.16% higher than the traditional damage function model. It can be seen that the model proposed in this patent is better than Resnet in all noise rates. The recognition accuracy of the method for label noise is given. Obviously, the method proposed by us has excellent robustness under different levels of label noise. Even with 10% noisy labels, our improved Resnet model performs well and achieves satisfactory accuracy. Under normal circumstances, when the wavelet time-frequency image is mapped to feature points, it is difficult to clearly see the boundaries of the feature point distribution, but after training the model, the features become separable. The proposed model is easily segmented even at high LNR. This proves that the model has good fault diagnosis and feature extraction capabilities.

[0097] In order to more clearly demonstrate the advantages of the present invention, we compare the results of the fault diagnosis method of the embodiment of the present invention with the results of the existing methods of multi-scale convolutional neural network-bidirectional long short term memory neural network (MSCNN-BiLSTM), variational mode decomposition-deep convolutional neural networks (VMD-DCNNs), classification adversarial autoencoders (Categorical Adversarial Autoencoders, CatAEE) and convolutional neural network-long short term memory neural network (CNN-LST) from the perspective of average test accuracy in a noisy environment. The results are shown in Table 3.

[0098] Table 3 Comparison of the average accuracy of the present invention (WT-IResNet) with other methods

[0099]

[0100] It can be seen from Table 3 that, compared with the existing methods, the fault diagnosis accuracy of the present invention is very competitive under different signal-to-noise ratios. Even when the signal-to-noise ratio is -4db, the fault diagnosis accuracy of the present invention still reaches more than 90%.

[0101] In summary, the fault diagnosis method based on wavelet transform and improved deep residual network in the embodiment of the present invention uses wavelet transform to complete the time-frequency feature extraction required for fault diagnosis, which not only mines out the rich information hidden in the original diagnostic signal, but also reduces the dependence on diagnostic experience and feature selection, and then uses the improved convolutional neural network to realize the fault diagnosis of rolling bearings under noisy environment and noise labels. A new loss function containing two adjustable parameters is used to replace the original loss function, so as to overcome the influence of human labeling errors on the final diagnosis results. It not only overcomes the disadvantage of traditional neural network technology that relies heavily on labels, frees people from the heavy manual labeling labor, greatly saves labor costs, but also significantly improves the accuracy and effectiveness of fault diagnosis. The present invention provides a new, feasible and effective way to solve the problem of rolling bearing fault diagnosis, which can be widely used in fault diagnosis in the fields of machinery, chemical industry, aerospace, etc. In addition, the new global singular value decomposition adaptive strategy strengthens the feature mapping and classification and can prevent the overfitting of the overall structure. Under noisy labels and environments, the performance of the method proposed by us is satisfactory.

[0102] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method, characterized in that: It includes the following steps: S1, dividing the collected vibration signal of the rolling bearing into a plurality of data segments connected one after another, and converting the collected vibration signal into a wavelet time-frequency image using a wavelet transform formula; S2, randomly divide and normalize all wavelet time-frequency images to obtain training sets, validation sets and test sets; S3. Determine the number of incorrectly labeled samples in the training set according to predetermined test indicators, where the test indicators include test accuracy and actual engineering experience; S4. Construct a deep residual network, which includes a convolution layer, a singular value decomposition pooling layer and a singular value decomposition adaptive layer. The singular value decomposition pooling layer is used to obtain the singular values ​​of multiple convolution kernels using the singular value decomposition method, and construct a singular value matrix of multiple convolution kernels as the output matrix of the singular value decomposition pooling layer. The singular value decomposition adaptive layer is used to calculate the singular value of each wavelet time-frequency image using the singular value decomposition method, and concatenate the singular values ​​of all wavelet time-frequency images into a vector as the input vector of the deep residual network softmax layer; S5. Add a loss function and an activation function to the deep residual network. The expression of the loss function is as follows: Among them, t1 and t2 are both constants, 0≤t1<1,t2>1, z j represents the input of the softmax layer, [·] + =max{·,0},γ t (z) by Calculate, N is the total number of training samples, x is the sample, i is the number of samples, c is the total number of fault categories, j is a constant of 1, 2, ... c; The activation function is shown below: S6. Optimize the learning rate of the deep residual network and use a dynamic learning rate instead of the commonly used static learning rate so that the learning rate gradually decreases as the training progresses. The formula for the dynamic learning rate is as follows: Among them, μ is the current learning rate; μ min is the minimum learning rate; μ0 is the initial learning rate; δ is the decay factor; gs is the current training step number; ds is the ratio of the number of samples in the batch to the number of samples in the entire training set; S7, using the training set to train the deep residual network, and using the verification set to verify the deep residual network of step S5, to obtain a trained fault diagnosis model with the best diagnostic effect verified by the verification set; S8. Use the test set as the input of the optimal fault diagnosis model obtained in step S6, and output the final fault diagnosis result.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that: The wavelet transform formula in step S1 is as follows: Among them, WT x (a,τ) is the wavelet transform coefficient, x(t) is the collected vibration signal, ψ a,τ (t) is the wavelet basis function, a is the scale factor, τ is the translation, specifically, the wavelet basis function is the Morlet wavelet, and the center frequency and bandwidth are both 3; t is the time for collecting the vibration signal.

3. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S1, a test platform is used to collect the vibration signal of the rolling bearing, and the test platform includes a motor, a torque sensor, a power meter and an electronic control device. The test platform includes a motor, a torque sensor, and a power meter that are respectively communicated with the electronic control device and controlled by the electronic control device; an acceleration sensor is used to collect the acceleration signal of the rolling bearing during operation, and the acceleration sensor is arranged along the axial or radial direction of the rotating axis of the rolling bearing.

4. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S1, Gaussian white noise is added to the collected vibration signal to simulate a strong noise environment. The Gaussian white noise is divided into 8 levels according to the signal-to-noise ratio, which are -4, -2, 0, 2, 4, 6, 8, and 10 from low to high.

5. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S2, a plurality of wavelet time-frequency images are randomly divided into a training set, a validation set and a test set as samples according to a set division ratio, and normalized respectively.

Images