Rolling bearing fault diagnosis method and system based on variational mode extraction

Abstract

The invention discloses a rolling bearing fault diagnosis method and system based on variational mode extraction, and the method comprises the following steps: collecting acceleration data corresponding to a rolling bearing fault, and dividing the acceleration data into a training sample and a test sample through overlapping sampling; designing a variational modal extraction module by using different parameters, and performing modal matrix extraction on the training sample and the test sample; constructing a deep convolutional fault diagnosis network model, inputting the modal matrix of the training sample into the deep convolutional fault diagnosis network model for training, and inputting the test sample into the trained deep convolutional fault diagnosis network model to realize bearing fault diagnosis; and carrying out interpretability analysis on the constructed deep convolutional fault diagnosis network model by using a gradient weighted class activation mapping method, selecting a modal component most concerned by the deep convolutional fault diagnosis network model from the modal matrix, carrying out envelope spectrum analysis, and verifying the interpretability of the model in combination with a theoretical fault characteristic frequency.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault diagnosis, and in particular to a rolling bearing fault diagnosis method and system based on variational mode extraction. Background Art

[0002] Rotating machinery is an important industrial equipment, widely used in wind power, hydropower, aviation and other fields. The main feature of this type of equipment is that it has rotating parts, which generate power and output it through mechanical torque. Bearings are one of the key components of rotating machinery, and their working condition directly affects the working efficiency and life of the entire system. During long-term operation, bearings will experience wear, damage, rolling element deformation and wear, so the diagnosis and prevention of bearing failures are crucial.

[0003] Fault diagnosis based on deep learning is a method of using deep neural networks to perform fault diagnosis on rotating mechanical equipment. Compared with traditional machine learning algorithms, deep learning can automatically extract features from a large amount of raw data, and realize the abstraction and representation of high-level features through multiple hidden layers, so it has stronger pattern recognition and prediction capabilities. However, in the past, data-driven deep learning fault diagnosis algorithms were often "black box" models, and users could not understand their internal operating principles and results, nor could they provide reliable explanations, which made it difficult for various deep learning models to be applied in actual production environments. In order to improve the interpretability of fault diagnosis of rotating mechanical equipment based on deep learning, the present invention discloses a rolling bearing fault diagnosis method and system based on variational mode extraction, which obtains a modal matrix containing typical fault features through variational mode extraction, combines a convolutional deep network to realize effective identification of bearing faults, and uses gradient-like activation mapping technology and envelope spectrum to perform interpretability analysis on the network. Summary of the invention

[0004] In response to the problems and needs raised above, this solution proposes a rolling bearing fault diagnosis method and system based on variational mode extraction. Due to the adoption of the following technical features, it can achieve the above technical objectives and bring about many other technical effects.

[0005] An object of the present invention is to provide a rolling bearing fault diagnosis method based on variational mode extraction, comprising the following steps: S10: collecting acceleration data corresponding to the rolling bearing fault, and dividing the acceleration data into training samples and test samples through overlapping sampling; S20: Design a variational mode extraction module using different parameters to extract the mode matrix of training samples and test samples; S30: construct a deep convolutional fault diagnosis network model, input the modal matrix of the training sample into the deep convolutional fault diagnosis network model for training, and input the test sample into the trained deep convolutional fault diagnosis network model to realize bearing fault diagnosis; S40: Use the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model. Select the modal component with the highest importance weight of the deep convolutional fault diagnosis network model from the modal matrix and perform envelope spectrum analysis. Combined with the theoretical fault characteristic frequency, the interpretability of the model is verified.

[0006] In addition, the rolling bearing fault diagnosis method based on variational mode extraction according to the present invention may also have the following technical features: In an example of the present invention, in step S20, designing a variational mode extraction module using different parameters includes: constructing a constrained variational model, the expression of which is as follows: , where f ( t ) is the input signal; u d ( t ) is the expected mode; f r ( t ) is the residual signal; δ is the Dirac distribution; ω d is the approximate center frequency; α is the quadratic penalty term, β ( t ) is the impulse response of the filter, j A plural unit.

[0007] In an example of the present invention, in step S20, a modal matrix is ​​extracted from the training sample and the test sample, wherein the modal matrix is ​​expressed as follows: , where i Represents the sample number, n represents the modal component length, m Represents the number of variational mode extractors.

[0008] In an example of the present invention, in the step S30, the deep convolutional fault diagnosis network model includes: a plurality of feature extraction modules connected in series and a fully connected layer, wherein the plurality of feature extraction modules are configured to perform multiple feature extractions on the modal matrix obtained by the variational modal extraction module; the fully connected layer is configured to implement discrimination and classification of different fault categories through a fully connected approach for the diagnostic results output by the feature extraction module.

[0009] In an example of the present invention, in the step S30, the feature extraction module includes: a two-dimensional convolution layer, a feature normalization layer, an activation function layer and a pooling layer connected in series, wherein the two-dimensional convolution layer is configured to extract local features in the input signal and capture key patterns and features related to rolling bearing faults through convolution operations; the feature normalization layer is configured to standardize the features output by the convolution layer; the activation function layer is configured to introduce nonlinear transformations to enhance the network expression capability; and the pooling layer is configured to downsample the features.

[0010] In an example of the present invention, in the step S30, it further includes: updating the deep convolutional fault diagnosis network model by a cross entropy loss function, and the expression of the cross entropy loss function is as follows: , where y i is the true label, N is the number of fault types, SoftMax The output is mapped to the (0,1) interval and normalized.

[0011] In an example of the present invention, in step S40, using the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model includes the following steps: S41: Prediction category for model output y c , use this output for back propagation to calculate the gradient information of the feature map of the last convolutional layer ,in, A k For the k feature maps; S42: Perform global average pooling on the gradient information of the feature map and calculate the importance weight α of each feature map.

[0012] In one example of the present invention, in step S42, the expression for calculating the importance weight of each feature map is: , where Z is the spatial size of the feature map; Indicates k The position in the feature map ( i , j ) value; y c It represents the predicted classification result; c represents the index of the predicted class.

[0013] In one example of the present invention, using the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model also includes the following steps: S43: The average gradient of the feature map of each channel is used as a weight to multiply the feature map of each channel and superimpose to obtain a class activation map. The negative values ​​are removed by the activation function Relu function to eliminate the influence of other categories. The formula for calculating the class activation map is: , where Importance weight for each feature map; A k That is k A feature map.

[0014] Another object of the present invention is to provide a fault diagnosis system for a rolling bearing fault diagnosis method based on variational mode extraction as described above, comprising: A data acquisition unit is configured to acquire acceleration data corresponding to the bearing fault, and divide the acceleration data into training samples and test samples by overlapping sampling; A modal matrix extraction unit is configured to design a variational modal extraction module using different parameters to extract modal matrices from training samples and test samples; A network model diagnosis unit is configured to construct a deep convolution fault diagnosis network model, input the modal matrix of the training sample into the deep convolution fault diagnosis network model for training, and input the test sample into the trained deep convolution fault diagnosis network model to realize bearing fault diagnosis; The network model verification unit is configured to use the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model, select the modal component with the highest importance weight of the deep convolutional fault diagnosis network model from the modal matrix and perform envelope spectrum analysis, and verify the interpretability of the model in combination with the theoretical fault characteristic frequency.

[0015] The best embodiment for carrying out the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0017] Figure 1 A logic diagram of a rolling bearing fault diagnosis method based on variational mode extraction according to an embodiment of the method of the present invention; Figure 2 A schematic diagram of a variational mode extraction module of an embodiment of the method of the present invention; Figure 3Schematic diagram of the structure of a deep convolutional fault diagnosis network model according to an embodiment of the present invention; Figure 4 A schematic diagram of the overall structure of a convolutional network for fusion variational mode extraction according to an embodiment of the method of the present invention; Figure 5 A time domain waveform diagram of a rotor fault data sample of a CWRU bearing according to an embodiment of the method of the present invention; Figure 6 The confusion matrix of the model of the method embodiment of the present invention on the test set; Figure 7 A model gradient weighted class activation map of an embodiment of the method of the present invention; Figure 8 This is an envelope spectrum analysis diagram of the 9th component of the test sample of the method embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Unless otherwise defined, the technical terms or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not necessarily indicate a quantity limitation. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Many intelligent diagnosis systems use deep learning technology, but deep learning models are essentially black box models, and their judgment results are difficult to explain, which makes the diagnosis lack credibility. Rotating machinery is a complex system, and its fault diagnosis is also a relatively complex task. In order to enhance users' trust in deep learning models, improving interpretability is essential.

[0021] According to a first aspect of the present invention, a rolling bearing fault diagnosis method based on variational mode extraction is provided. Figures 1 to 4 As shown, the following steps are included: S10: collecting acceleration data corresponding to the rolling bearing fault, and dividing the acceleration data into training samples and test samples through overlapping sampling; S20: Design a variational mode extraction module using different parameters to extract the mode matrix of training samples and test samples; S30: construct a deep convolutional fault diagnosis network model, input the modal matrix of the training sample into the deep convolutional fault diagnosis network model for training, and input the test sample into the trained deep convolutional fault diagnosis network model to realize bearing fault diagnosis; S40: Use the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model. Select the modal component with the highest importance weight of the deep convolutional fault diagnosis network model from the modal matrix and perform envelope spectrum analysis. Combined with the theoretical fault characteristic frequency, the interpretability of the model is verified.

[0022] This method uses variational mode extraction technology to extract the mode matrix from the original signal, and extracts the mode containing fault information from the mode matrix through a convolutional neural network. A deep convolution fault diagnosis network model is constructed, which has the advantages of signal processing and convolutional networks. It can integrate signal processing knowledge into the deep learning model to achieve high-precision fault diagnosis of rotating machinery equipment, and the network is interpretably analyzed through the gradient weighted class activation map. Compared with traditional intelligent diagnosis methods, this network has strong interpretability while integrating signal processing knowledge, which is convenient for engineering applications.

[0023] This method extracts the modal matrix from the collected vibration signal through variational mode extraction, and uses convolutional neural network to automatically extract the mode containing fault information to achieve accurate and efficient fault diagnosis. The model is interpretable through gradient class activation mapping technology and envelope spectrum, which has a good guiding value for improving the interpretability and credibility of rotating machinery fault diagnosis.

[0024] In one example of the present invention, in step S20, as Figure 2 As shown, designing a variational mode extraction module using different parameters includes: constructing a constrained variational model, which is expressed as follows: , where f ( t ) is the input signal; u d ( t ) is the expected mode; f r ( t ) is the residual signal; δ is the Dirac distribution; ω d is the approximate center frequency; α is the quadratic penalty term, β ( t ) is the impulse response of the filter, j A plural unit.

[0025] In an example of the present invention, in step S20, a modal matrix is ​​extracted from the training sample and the test sample, wherein the modal matrix is ​​expressed as follows: , where i Represents the sample number, n represents the modal component length, m Represents the number of variational mode extractors.

[0026] In one example of the present invention, in step S30, Figure 3 As shown, in an example of the present invention, in the step S30, the deep convolutional fault diagnosis network model includes: a plurality of feature extraction modules and a fully connected layer connected in series, wherein the plurality of feature extraction modules are configured to perform multiple feature extractions on the modal matrix obtained by the variational modal extraction module; the fully connected layer is configured to use the diagnosis results output by the feature extraction module to achieve discrimination and classification of different fault categories through a fully connected manner.

[0027] Among them, the feature extraction module includes: a two-dimensional convolution layer, a feature normalization layer, an activation function layer and a pooling layer connected in series in sequence, wherein the two-dimensional convolution layer is configured to extract local features in the input signal and capture key patterns and features related to rolling bearing faults through convolution operations; the feature normalization layer is configured to standardize the features output by the convolution layer; the activation function layer is configured to introduce nonlinear transformations to enhance the network's expression capabilities; and the pooling layer is configured to downsample the features.

[0028] Specifically, the two-dimensional convolution layer is configured to extract local features in the input signal and capture key patterns and features related to rolling bearing faults through convolution operations; the feature normalization layer is configured to standardize the features output by the convolution layer to eliminate the scale differences between different features and improve the training efficiency and convergence speed of the model; the Relu activation layer is configured to introduce nonlinear transformations to enhance the expression ability of the network so that it can better fit complex fault modes; the pooling layer is configured to downsample the features to reduce the feature dimension and computational burden while retaining the main feature information and enhancing the anti-interference ability of the model; the fully connected layer is configured to map the high-dimensional features output by the pooling layer to the final diagnostic results and achieve discrimination and classification of different fault categories through a fully connected approach.

[0029] In one example of the present invention, the expression of the two-dimensional convolutional layer is as follows: , where represents the feature vector, represents the modal matrix of the input, i is the sample number, W , b Represent the weight and bias vector of the calculation process respectively, realizing the nonlinear transformation extraction of features. R (·) represents the activation function Relu.

[0030] In an example of the present invention, in the step S30, it further includes: updating the deep convolutional fault diagnosis network model by a cross entropy loss function, and the expression of the cross entropy loss function is as follows: , where y i is the true label, N is the number of fault types, SoftMax The output is mapped to the (0,1) interval and normalized.

[0031] In an example of the present invention, in step S40, using the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model includes the following steps: S41: Prediction category for model output y c , use this output for back propagation to calculate the gradient information of the feature map of the last convolutional layer ,in, A k That is k feature maps; S42: Perform global average pooling on the gradient information of the feature map and calculate the importance weight α of each feature map.

[0032] The employed gradient-weighted class activation map based interpretability technique exploits the flow of gradients backpropagated from output neurons to convolutional layers to understand model predictions.

[0033] In one example of the present invention, in step S42, the expression for calculating the importance weight of each feature map is: In the formula, Z is the spatial size of the feature map (i.e., the total number of all spatial locations in the feature map); represents the position in the kth feature map ( i , j ) value; y c It represents the predicted classification result; c represents the index of the predicted class.

[0034] In one example of the present invention, using the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model also includes the following steps: S43: The average gradient of the feature map of each channel is used as a weight to multiply the feature map of each channel and superimpose to obtain a class activation map. The negative values ​​are removed by the activation function Relu function to eliminate the influence of other categories. The formula for calculating the class activation map is: , where Importance weight for each feature map; A k That is k A feature map.

[0035] Finally, the class activation map is superimposed on the input modal matrix in the form of a heat map to obtain a visualization showing which component of the modal matrix the model focuses on when making fault classification decisions.

[0036] It should be noted that the theoretical fault characteristic frequency of various faults can be calculated based on the relevant parameters of the bearing and the speed of the rotating mechanical equipment. The calculation formula is as follows: , where , and are the characteristic frequencies of outer ring fault, inner ring fault and roller element fault respectively, D is the bearing diameter, is the diameter of the roller, is the contact angle, n is the number of rolling elements, S is the bearing speed.

[0037] According to the second aspect of the present invention, a fault diagnosis system of a rolling bearing fault diagnosis method based on variational mode extraction as described above comprises: A data acquisition unit is configured to acquire acceleration data corresponding to the bearing fault, and divide the acceleration data into training samples and test samples by overlapping sampling; A modal matrix extraction unit is configured to design a variational modal extraction module using different parameters to extract modal matrices from training samples and test samples; A network model diagnosis unit is configured to construct a deep convolution fault diagnosis network model, input the modal matrix of the training sample into the deep convolution fault diagnosis network model for training, and input the test sample into the trained deep convolution fault diagnosis network model to realize bearing fault diagnosis; The network model verification unit is configured to use the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model, select the modal component with the highest importance weight of the deep convolutional fault diagnosis network model from the modal matrix and perform envelope spectrum analysis, and verify the interpretability of the model in combination with the theoretical fault characteristic frequency.

[0038] The system uses variational mode extraction technology to extract the mode matrix from the original signal, and extracts the mode containing fault information from the mode matrix through a convolutional neural network. A convolutional network model that integrates variational mode extraction is constructed, which has the advantages of signal processing and convolutional networks. It can integrate signal processing knowledge into the deep learning model to achieve high-precision fault diagnosis of rotating machinery equipment, and the network is interpretably analyzed through the gradient weighted class activation map. Compared with traditional intelligent diagnosis methods, this network has strong interpretability while integrating signal processing knowledge, which is convenient for engineering applications. Specific examples

[0039] The collected fault signal data is divided by overlapping sampling and equal interval segmentation. The length of each data sample is 2048. There is an overlap between each signal segment and the next signal segment. The length of the overlapping part is 28. There are 1000 samples of each type. The ratio of training set to test set is 7:3. The overlapping data can achieve data enhancement and increase the number of training samples to ensure the balance of fault types.

[0040] A convolutional neural network model integrating variational mode extraction is established, referred to as VME-CNN. It is mainly composed of one variational mode extraction module, five feature extraction modules and one classification module. Among them, five feature extraction modules and one classification module constitute a deep convolutional fault diagnosis network model.

[0041] The training samples are used to train the VME-CNN model to realize the fault feature learning in the variational modal matrix, and the test samples are input into the trained VME-CNN network model to realize the automatic identification and intelligent diagnosis of the operating status of the wind turbine.

[0042] The modal components that the model focuses on are located through the gradient class activation map, and envelope spectrum analysis is performed on them to prove whether the features extracted by the network contain fault feature information.

[0043] Furthermore, overlapping sampling is used when dividing training samples and test samples. That is, when collecting training samples from the original signal, there is overlap between each signal segment and the next signal segment. For the test samples, there is no overlap when collecting. In order to facilitate the training of the convolutional neural network, each signal segment x is normalized. The normalization formula is as follows: In the formula, x min is the minimum value in the signal, x max is the maximum value in the signal.

[0044] Further, if Figure 4 As shown in the figure, the established VME-CNN model consists of a variational mode extraction module to extract the modal matrix of the original signal, five feature extraction modules and a classification module to realize feature extraction and fault classification.

[0045] The variational mode extraction module constructs 16 different mode extraction layers by changing parameters, starting from a length of n Extract the signal of size 16* n The modal matrix of the variational mode extraction module is initialized using the following formula; , where f s is the sampling frequency, ω d represents the approximate center frequency, α is the penalty coefficient.

[0046] The feature extraction module consists of a two-dimensional convolution layer, a feature normalization layer, an activation function layer and a maximum pooling layer, and is mainly used for more in-depth processing of the modal matrix of the signal to obtain more representative fault feature information.

[0047] Furthermore, the classification module is a fully connected layer, and the Softmax function used is specifically expressed as follows: , where p i Indicates i The probability that a class fault sample is correctly identified, z i Indicates i The output value of the network node, Cis the number of categories of fault samples.

[0048] Furthermore, the VME-CNN network model parameters are initialized, including the maximum number of training rounds. M =50, initialize learning rate lr =0.001, initialize batch size S =32, and the network parameters are iteratively updated through the Adam algorithm.

[0049] Specific implementation case: The bearing data of Case Western Reserve University (CWRU) is used to verify the effectiveness of the present invention. The CWRU bearing fault data set contains bearing fault data at the drive end and the fan end. The bearing speed range is 1730rpm-1797rpm, and the sampling frequency of the vibration data is 12kHz and 48kHz. The bearing fault is made by electrospark machining, including single-point faults of varying degrees at three positions: the inner ring, the rolling element, and the outer ring of the bearing. The fault sizes are 0.007inch, 0.014inch, and 0.021inch, respectively. The test sample of the present invention uses drive end data, the bearing model is 6205-2RS JEM SKF deep groove ball bearing, and 10 fault state vibration signals including normal state are selected, with a sampling frequency of 12kHz. Data division is performed by overlapping sampling and equal interval segmentation. The length of each data sample is 2048, and there is an overlap between each signal and the next signal. The length of the overlapping part is 28, and there are 1000 samples of each type. The ratio of the training set and the test set is 7:3. Figure 5 This is the time domain waveform of a rotor fault data sample.

[0050] In order to prove the effectiveness and interpretability of the proposed method in rotating machinery fault diagnosis, the trained deep learning model is tested using test data to obtain the diagnosis results and draw the confusion matrix, as shown in Figure 2. Figure 6 As shown in Figure 2, it can be clearly seen that the proposed VME-CNN model has achieved 100% diagnostic accuracy. In addition, taking the outer ring fault as an example, it is input into the model and the model gradient weighted class activation map is plotted, as shown in Figure 2. Figure 7 As shown in , the greater the importance weight of the modal component, the greater the contribution of the component to the fault diagnosis result, and the model pays more attention to this part of the data. Figure 8 As shown, the 9th component is selected for envelope spectrum analysis, and the vertical line marks the theoretical fault characteristic frequency calculated according to the formula. After comparison, it can be found that the proposed VME-CNN model can accurately extract the modal components containing obvious fault characteristic information from the modal matrix, and has strong interpretability. This shows that the VME-CNN model disclosed in the present invention has obvious diagnostic advantages compared with the traditional deep learning model, can realize the intelligent identification and reliable diagnosis of rotating machinery equipment faults, and has good interpretability.

[0051] The exemplary implementation of the rolling bearing fault diagnosis method and system based on variational modal extraction proposed in the present invention is described in detail above with reference to the preferred embodiments. However, those skilled in the art will appreciate that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the protection scope of the present invention, which is determined by the attached claims.

Claims

1. A rolling bearing fault diagnosis method based on variational mode extraction, characterized in that: The steps include: S10: collecting acceleration data corresponding to the rolling bearing fault, and dividing the acceleration data into training samples and test samples through overlapping sampling; S20: Design a variational mode extraction module using different parameters to extract the mode matrix of training samples and test samples; S30: construct a deep convolutional fault diagnosis network model, input the modal matrix of the training sample into the deep convolutional fault diagnosis network model for training, and input the test sample into the trained deep convolutional fault diagnosis network model to realize bearing fault diagnosis; S40: Use the gradient weighted class activation mapping method to perform interpretability analysis on the deep convolutional fault diagnosis network model. Select the modal component with the highest importance weight of the deep convolutional fault diagnosis network model from the modal matrix and perform envelope spectrum analysis. Combined with the theoretical fault characteristic frequency, the interpretability of the model is verified.

2. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 1 is characterized in that: In the step S20, designing a variational mode extraction module using different parameters includes: constructing a constrained variational model, the expression of which is as follows: , where f ( t ) is the input signal; u d ( t ) is the expected mode; f r ( t ) is the residual signal; δ is the Dirac distribution; ω d is the approximate center frequency; α is the quadratic penalty term, β ( t ) is the impulse response of the filter, j A plural unit.

3. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 2 is characterized in that: In step S20, the modal matrix is ​​extracted from the training sample and the test sample, wherein the expression of the modal matrix is ​​as follows: , where i Represents the sample number, n represents the modal component length, m Represents the number of variational mode extractors.

4. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 1 is characterized in that: In step S30, the deep convolutional fault diagnosis network model includes: a plurality of feature extraction modules and a fully connected layer connected in series, wherein the plurality of feature extraction modules are configured to perform multiple feature extractions on the modal matrix obtained by the variational modal extraction module; and the fully connected layer is configured to use a fully connected method to implement discrimination and classification of different fault categories using the diagnostic results output by the feature extraction module.

5. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 4 is characterized in that: In step S30, the feature extraction module includes: a two-dimensional convolution layer, a feature normalization layer, an activation function layer and a pooling layer connected in series, wherein the two-dimensional convolution layer is configured to extract local features in the input signal and capture key patterns and features related to rolling bearing faults through convolution operations; the feature normalization layer is configured to standardize the features output by the convolution layer; the activation function layer is configured to introduce nonlinear transformations to enhance network expression capabilities; and the pooling layer is configured to downsample the features.

6. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 1 is characterized in that: In the step S30, it also includes: updating the deep convolutional fault diagnosis network model by a cross entropy loss function, and the expression of the cross entropy loss function is as follows: , where y i is the true label, N is the number of fault types, SoftMax The output is mapped to the (0,1) interval and normalized.

7. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 1 is characterized in that: In step S40, using the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model includes the following steps: S41: Prediction category for model output y c , use this output for back propagation to calculate the gradient information of the feature map of the last convolutional layer ,in, A k For the k feature maps; S42: Perform global average pooling on the gradient information of the feature map and calculate the importance weight α of each feature map.

8. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 7 is characterized in that: In step S42, the expression for calculating the importance weight of each feature map is: , where Z is the spatial size of the feature map; Indicates k The position in the feature map ( i , j ) value; y c It represents the predicted classification result; c represents the index of the predicted class.

9. The rolling bearing fault diagnosis method based on variational mode extraction according to claim 7 is characterized in that: The interpretability analysis of the constructed deep convolutional fault diagnosis network model using the gradient weighted class activation mapping method also includes the following steps: S43: The average gradient of the feature map of each channel is used as a weight to multiply the feature map of each channel and superimpose to obtain a class activation map. The negative values ​​are removed by the activation function Relu function to eliminate the influence of other categories. The formula for calculating the class activation map is: , where Importance weight for each feature map; A k That is k A feature map.

10. A fault diagnosis system using the rolling bearing fault diagnosis method based on variational mode extraction as described in any one of claims 1 to 9, characterized in that: include: A data acquisition unit is configured to acquire acceleration data corresponding to the bearing fault, and divide the acceleration data into training samples and test samples by overlapping sampling; A modal matrix extraction unit is configured to design a variational modal extraction module using different parameters to extract modal matrices from training samples and test samples; A network model diagnosis unit is configured to construct a deep convolution fault diagnosis network model, input the modal matrix of the training sample into the deep convolution fault diagnosis network model for training, and input the test sample into the trained deep convolution fault diagnosis network model to realize bearing fault diagnosis; The network model verification unit is configured to use the gradient weighted class activation mapping method to perform interpretability analysis on the constructed deep convolutional fault diagnosis network model, select the modal component with the highest importance weight of the deep convolutional fault diagnosis network model from the modal matrix and perform envelope spectrum analysis, and verify the interpretability of the model in combination with the theoretical fault characteristic frequency.

Images