CN116625689B - Rolling bearing fault diagnosis method and system based on SMDER - Google Patents

Abstract

The invention discloses a rolling bearing fault diagnosis method and system based on SMDER, wherein the method comprises the following steps: the bearing vibration signal is subjected to continuous wavelet transformation to obtain the time-frequency domain characteristics of the original signal, and then the time-frequency domain characteristics are input into the sharing module for processing; and inputting the features extracted by the sharing module into a super-feature extractor for feature extraction, and then using a classifier for classification processing to realize bearing fault diagnosis under the condition of incremental learning. The method can enable the network model to have a good identification effect on both new and old data after learning the new data.

Description

Rolling bearing fault diagnosis method and system based on SMDER

Technical field

The present invention relates to the technical field of bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis method and system based on SMDER.

Background technique

Common rolling bearing fault diagnosis methods include: vibration analysis method, oil film resistance monitoring method, acoustic emission method, temperature monitoring method, oil monitoring method, optical fiber signal monitoring method, voltage and current detection method, etc. Fault diagnosis of rolling bearings based on temperature signals is easily affected by ambient temperature, and only has a strong diagnostic effect on features that appear later in the fault. Therefore, fault diagnosis of rolling bearings based on vibration signals can avoid the interference of ambient temperature and has higher research value. Methods for fault diagnosis of rolling bearings based on vibration signals include: signal processing-based methods and deep learning-based methods.

In the existing technology, the MTCN network model improves the accuracy of rolling bearing fault diagnosis under various speeds and variable load conditions. However, under the actual operating status of the bearing, fault data is gradually generated, and the MTCN network model is a batch learning model. When new bearing fault data arrives, training the traditional batch learning model using only new data will face the problem of catastrophic forgetting. , resulting in the model only having better recognition effect on new data, but poor recognition effect on old data.

Contents of the invention

The technical problem to be solved by the present invention is how to provide a rolling bearing fault diagnosis method that can enable the network model to learn new data and simultaneously have a good recognition effect on old and new data.

In order to solve the above technical problems, the technical solution adopted by the present invention is: a rolling bearing fault diagnosis method based on SMDER, which includes the following steps:

Pass the bearing vibration signal through continuous wavelet transform to obtain the time-frequency domain characteristics of the original signal, and then input the time-frequency domain characteristics into the shared module for processing;

Input the features extracted by the shared module into the super feature extractor for feature extraction and then use the classifier for classification processing to achieve bearing fault diagnosis under incremental learning conditions.

A further technical solution is that the shared module first inputs the wavelet time-frequency map to the first convolution layer of the shared module for convolution processing, then performs batch normalization BatchNorm, and uses the ReLU activation function to process, and then The obtained feature vector is input to the second convolution layer, and finally batch normalization is performed. After the shared module operation is completed, the obtained features are input to the feature extractor of each task for processing according to the new tasks. .

A further technical solution is that the super feature extractor includes multiple feature extractors, each feature extractor corresponding to an incremental learning task, in which the old feature extractor is used to extract features of existing data. After the new class arrives, Freeze the existing feature extractor parameters, and add a new feature extractor to extract the features of the new class. At this time, the combination of all feature extractors is a super feature extractor, and then the features extracted by all feature extractors are spliced. That is, the features extracted by the hyper-feature extractor; the formula of the hyper-feature extractor is as shown in Equation (1):

z＝F _t (x)＝[F _t-1 (x), Y _t (x)] (1)

In the formula, x is the feature output by the shared module and is also the input of the super feature extractor F _t ; the super feature extractor is spliced by the super feature extractor F _t-1 of the previous step and the new feature extractor Y _t of the current step Composition; z is the feature extracted by the super feature extractor.

A further technical solution is to use a differentiable channel-level mask method to prune the feature extractor. The channel-level mask and features are learned together. After learning the mask, the mask is binarized and the The feature extractor performs pruning to obtain the pruned network; the channel-level mask is shown in Equation (2):

h′ _l ＝h _l ⊙σ(se _l ) (2)

where h' _l is the mask feature after pruning; h _l is the mask feature before pruning; ⊙ is the channel-level multiplication; σ is the sigmoid gating function, whose value range is [0,1]; s is Scaling factor; e _l is the parameter of the mask, which can be learned during training.

A further technical solution is: the loss function of the entire network is as shown in Equation (3):

In the formula, L is the network loss; is the training loss; λ _α is the hyperparameter of the auxiliary classifier, and the initial value is 0;/> is the auxiliary loss; λ _s is the hyperparameter that controls the size of the model; _LS is the sparse loss.

The further technical solution is: the old feature parameters of the classifier H _t are the parameters frozen in the H _t-1 step used, and the parameters of the new features are randomly initialized; the feature z extracted by the super feature extractor is input to the classifier H _t is used for classification and prediction, as shown in formula (4):

y＝argmax(Softmax(H _t (z))) (4)

In the formula, y is the final predicted classification result; argmax is the point set whose independent variable reaches the maximum value; Softmax is the activation function of the output layer, which converts the prediction result into a non-negative number; H _t is the classifier; z is the super feature Features extracted by the extractor.

The invention also discloses a SMDER-based rolling bearing fault diagnosis system, which includes:

The wavelet transformation module is used to transform the bearing vibration signal through continuous wavelet transformation to obtain the time-frequency domain characteristic map of the original signal;

The shared module is used to extract features from the input time-frequency domain feature map;

Super feature extractor, including multiple feature extractors, each feature extractor corresponds to an incremental learning task, in which the old feature extractor is used to extract features of existing data. After the new class arrives, the parameters of the existing feature extractor are frozen. , and add a new feature extractor to extract the features of the new class. At this time, the combination of all feature extractors is a super feature extractor, and then the features extracted by all feature extractors are spliced together, which is the super feature extractor extraction. Characteristics;

The classifier is used to classify the features extracted by the super feature extractor to realize bearing fault diagnosis under incremental learning conditions.

The beneficial effect of adopting the above technical solution is that the method inputs the time-frequency domain characteristics of the bearing vibration signal obtained through wavelet transformation into the sharing module to share old and new data and fault characteristics. Then the extracted features are input to the super feature extractor for feature extraction and classification to achieve the purpose of bearing fault diagnosis under incremental learning conditions. Experimental results show that among the existing incremental learning models, the SMDER model has a higher recognition rate for new fault types, which solves the problem of catastrophic forgetting to a certain extent and enables the network model to learn new data while simultaneously recognizing both old and new data. Better recognition effect.

Description of the drawings

The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Figure 1 is a structural diagram of the DER network model in the method according to the embodiment of the present invention;

Figure 2 is an original processing flow diagram of the method according to the embodiment of the present invention;

Figure 3 is a time domain signal diagram of the bearing in the embodiment of the present invention;

Figure 4 is the bearing time-frequency domain signal after continuous wavelet transformation in the embodiment of the present invention;

Figure 5 is a functional block diagram of a shared module in an embodiment of the present invention;

Figure 6 is a physical diagram of the bearing test bench of Case Western Reserve University in an embodiment of the present invention;

Figure 7 is a test set accuracy curve chart after completing the new tasks in the embodiment of the present invention;

Figure 8 is a structural diagram of the ResNet10 network model in the embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention are clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Those skilled in the art can do so without departing from the connotation of the present invention. Similar generalizations are made, and therefore the present invention is not limited to the specific embodiments disclosed below.

The embodiment of the present invention discloses a rolling bearing fault diagnosis method based on SMDER, which includes the following steps:

Pass the bearing vibration signal through continuous wavelet transform to obtain the time-frequency domain characteristics of the original signal, and then input the time-frequency domain characteristics into the shared module for processing;

Input the features extracted by the shared module into the super feature extractor for feature extraction and then use the classifier for classification processing to achieve bearing fault diagnosis under incremental learning conditions.

The above method is explained in detail below with reference to relevant specific content:

Dynamically Expandable Representation (DER) is proposed for incremental learning to avoid "catastrophic forgetting". The network model is composed of a hyper-feature extractor and a classifier. The DER network model structure is shown in Figure 1.

hyperfeature extractor

The super feature extractor is composed of multiple feature extractors, and each feature extractor corresponds to an incremental learning task. The old feature extractor can extract features from existing data. After the new class arrives, the parameters of the existing feature extractor are frozen, and a new feature extractor is added to extract the features of the new class. At this time, the combination of all feature extractors is a super feature extractor, and then the features extracted by all feature extractors are spliced together, which is the feature extracted by the super feature extractor. The formula of the super feature extractor is shown in Equation (1).

z＝F _t (x)＝[F _t-1 (x), Y _t (x)] (1)

In the formula, x is the feature output by the shared module and is also the input of the super feature extractor F _t ; the super feature extractor is spliced by the super feature extractor F _t-1 of the previous step and the new feature extractor Y _t of the current step. Composition; z is the feature extracted by the super feature extractor.

channel level mask

Since a new feature extractor is added when a task is added, the model structure becomes larger and larger as tasks continue to be added. Therefore, a differentiable channel-level mask method is used to prune the feature extractor. The channel-level mask and features are learned together. After learning the mask, the mask is binarized and the feature extractor is pruned to obtain the pruned network. DER's channel-level mask is improved based on HAT. The mask of the entire channel is used instead of the mask of HAT's filter weight, thereby increasing the intensity of pruning. This channel-level mask is applied after the first convolutional layer of the new feature extractor. The channel-level mask is shown in Equation (2).

h′ _l ＝h _l ⊙σ(se _l ) (2)

where h' _l is the mask feature after pruning; h _l is the mask feature before pruning; ⊙ is the channel-level multiplication; σ is the sigmoid gating function, whose value range is [0,1]; s is Scaling factor; e _l is the parameter of the mask, which can be learned during training.

network loss

The loss of the entire network is set to the sum of cross-entropy loss, auxiliary loss and sparse loss. The training loss is cross-entropy loss. The purpose of adding auxiliary loss is to enable the network model to distinguish between old and new tasks, so that the network model has a certain discriminative ability. Reduce network parameters by adding sparse loss. Therefore, the loss function of the entire network is shown in Equation (3).

In the formula, L is the network loss; is the training loss; λ _α is the hyperparameter of the auxiliary classifier, and the initial value is 0;/> is the auxiliary loss; λ _s is the hyperparameter that controls the size of the model; _LS is the sparse loss.

Classifier

The old feature parameters of the classifier H _t are the parameters frozen in the H _t-1 step used, and the parameters of the new features are randomly initialized. The feature z extracted by the super feature extractor is input to the classifier H _t for classification and prediction, as shown in Equation (4).

y＝argmax(Softmax(H _t (z))) (4)

In the formula, y is the final predicted classification result; argmax is the point set whose independent variable reaches the maximum value; Softmax is the activation function of the output layer, which converts the prediction result into a non-negative number; H _t is the classifier; z is the super feature Features extracted by the extractor.

After learning the super features, the classifier is retrained using all previous data and the current new data at the current time step, and the balanced fine-tuning method is used to deal with the class imbalance problem.

Dynamically scalable network based on shared modules (SMDER)

On the basis of DER, a shared module is added, and a dynamically expandable network SMDER (Shared Module Dynamically Expandable Representation) based on shared modules is proposed. The overall structure of SMDER is shown in Figure 2.

As can be seen from Figure 2, SMDER first performs wavelet transformation on each new task sample to obtain a wavelet time-frequency diagram, and then inputs the wavelet time-frequency diagram obtained from each new task into the shared module to extract basic features. Starting from basic features, dynamic network expansion is performed for different tasks, and a feature extractor is added for each new task. Finally, the features of all feature extractors will be spliced and the classifier will be used to complete the classification. Specific steps are as follows:

First, the time-frequency domain features of the original signal are obtained through continuous wavelet transform, and then the time-frequency domain features are input to the shared module. Taking the vibration signal of an inner ring fault with a fault size of 0.014 inches as an example, the original time domain signal is shown in Figure 3, and the time-frequency domain signal obtained after continuous wavelet transformation is shown in Figure 4.

The time-frequency domain features extracted from Figure 4 can clearly see the frequency and appearance time of the signal, which provides a good feature basis for subsequent model expansion. Before inputting SMDER, delete the coordinate axes and energy bars that have nothing to do with the features in the time-frequency diagram to facilitate basic feature extraction by the network model.

Secondly, because the time-frequency domain features extracted from vibration signals are expressed in the form of pictures, and convolutional neural networks have special advantages in extracting picture features, a shared module was designed based on the convolutional neural network to extract time-frequency Domain features serve as the basic features for each task. This shared module consists of two convolutional layers, a normalization operation and an activation function, as shown in Figure 5.

The sharing module first inputs the wavelet time-frequency image into the first convolution layer of the sharing module. The convolution kernel size of the first convolution layer is 3*3, the number of channels is 64, and the step size is 1. Then perform batch normalization BatchNorm, and use the ReLU activation function to avoid the problems of gradient explosion and gradient disappearance, thereby alleviating the problem of overfitting and speeding up the convergence rate. The obtained feature vector is then input to the second convolution layer. The convolution kernel size of the second convolution layer is 3*3, the number of channels is 32, and the stride is 1. Finally, batch normalization BatchNorm is performed. After the shared module operation is completed, the obtained features will be input to the feature extractor of each task according to the new tasks. Since the basic features of image data are similar, this shared module is set to the lowest layer of the entire network model.

Correspondingly, embodiments of the present invention also disclose a rolling bearing fault diagnosis system based on SMDER, including:

The wavelet transformation module is used to transform the bearing vibration signal through continuous wavelet transformation to obtain the time-frequency domain characteristic map of the original signal;

The shared module is used to extract features from the input time-frequency domain feature map;

Super feature extractor, including multiple feature extractors, each feature extractor corresponds to an incremental learning task, in which the old feature extractor is used to extract features of existing data. After the new class arrives, the parameters of the existing feature extractor are frozen. , and add a new feature extractor to extract the features of the new class. At this time, the combination of all feature extractors is a super feature extractor, and then the features extracted by all feature extractors are spliced together, which is the super feature extractor extraction. Characteristics;

The classifier is used to classify the features extracted by the super feature extractor to realize bearing fault diagnosis under incremental learning conditions.

Experimental verification and analysis

The experimental environment and related training parameter settings are shown in Table 1.

Table 1 Experimental environment and related training parameter settings

The bearing failure accuracy mentioned in incremental learning is the average incremental accuracy, that is, the model obtained after training using the training set of the current task is tested on the test set of the current task and all previous tasks, and finally The obtained test accuracy of all tasks is averaged, which is the average incremental accuracy.

Experimental data set

The data set is the Case Western Reserve University (CWRU) bearing data set. The experimental bench is shown in Figure 6. The driving end bearing fault data is selected, and the sampling frequency is 12KHz. The bearing brand and model is SKF 6205-2RS, the rotation speed is 1750r/min, and the load is 2HP. According to the fault diameter and fault location of the bearing, it is divided into 9 categories, plus the data of non-faulty bearings, there are 10 categories in total. The details of the working conditions are shown in Table 2.

Table 2 Operating conditions used in the Case Western Reserve University bearing data set

In addition, in actual working conditions, it is detected that the data arrives in batches. Therefore, for incremental learning, the incremental sequence of tasks must be set to simulate the actual operating status. The sequence of tasks in this application is (5,8,2,6, 9,3,4,0,7,1).

Network parameter settings

The number of learning times of the shared module is 1, and the parameter settings of the shared module are shown in Table 3.

Table 3 Parameter settings of shared modules

The 18 in Resnet18 refers to the number of convolutional layers or fully connected layers being 18. The parameter settings of the feature extractor Resnet18 are shown in Table 4.

Table 4 Parameter settings of feature extractor Resnet18

Experimental results

The experimental process takes the i-th task as an example, where i = 1 to 10, that is, all bearing data categories to be learned: SMDER is trained using the training set of the i-th task, and the model after training uses the 1 to i The test set of the task is tested. The purpose of the test is to prove that the proposed SMDER model can avoid "catastrophic forgetting", that is, the model can learn new tasks while ensuring the recognition accuracy of old tasks. Under the condition that each new category is added, the accuracy curve of the test set after adding each task is shown in Figure 7.

As can be seen from Figure 7, the accuracy of SMDER in the test set of the first five tasks can reach 100%, and the accuracy of the test set of the subsequent five tasks remains above 89%. Experiments show that the average incremental accuracy of 10 tasks can reach 96.7%. Therefore, the SMDER network model has a higher accuracy and better effect in bearing fault diagnosis under incremental learning.

ablation experiment

In order to verify the impact of the convolution kernel scale of the shared module, the number of channels of the shared module, and the number of learning times of the shared module on the experimental results proposed in this application, an ablation experiment was performed here. The feature extractors used in this ablation experiment are all Resnet18.

The impact of the number of channels of shared modules on experimental results

SMDER's shared module has two convolutional layers. In order to compare the impact of shared modules with different channel numbers on bearing fault diagnosis accuracy, this application uses (16,16), (16,32), ( A total of 9 different channel numbers were tested. In this experiment, the convolution kernel scale of the shared module is (3*3), the step size of the convolution is 1, and the number of learning times of the shared module is 1. The experimental results are shown in Table 5.

Table 5 Diagnostic accuracy of shared modules with different channel numbers (Channels)

It can be seen from Table 5 that when the number of channels of the shared module is (64,32), the bearing fault diagnosis accuracy is the highest, reaching 96.70%; when the number of channels of the shared module is (32,64), the bearing fault diagnosis accuracy is 96.55 %. It can be seen that the number of shared module channels is 32/64, which plays an important role in the accuracy of bearing fault diagnosis.

The impact of the convolution kernel scale of the shared module on the experimental results

In order to compare the impact of shared modules of different convolution kernel sizes on the accuracy of rolling bearing fault diagnosis, a total of three different scales of convolutions (1*1), (3*3), and (5*5) were used for the two convolution layers. The experiments were carried out. At the same time, due to the aforementioned shared module channel number experiment, when the channel numbers of the two convolutional layers of the shared module are (64,32) and (32,64), the bearing fault diagnosis accuracy can reach more than 96%, and the difference is Not big. Therefore, the experiment uses these two channel numbers and three convolution kernels of different scales to conduct combined experiments. In this experiment, the step size of the convolution is 1, and the number of shared module learning times is 1. The experimental results are shown in Table 6.

Table 6 Diagnostic accuracy of shared modules with different convolution kernel scales (Size)

As can be seen from Table 6, when the convolution kernel scale is (1*1), the training time is the longest and the accuracy is the lowest. When the convolution kernel scale is (5*5), the training time is the shortest, but the accuracy is between the convolution kernel scale (1*1) and (5*5). When the convolution kernel scale is (3*3), the bearing fault diagnosis accuracy can reach the highest under the conditions of the two shared module channel numbers. Therefore, the optimal convolution kernel scale of the shared module is (3*3).

The impact of the number of learning times of shared modules on experimental results

In order to compare the impact of different learning times of the shared module on the accuracy of bearing fault diagnosis, experiments were conducted using a total of 5 different learning times of 1, 2, 3, 4, and 5 for the shared module. The number of shared module learning times refers to the number of new tasks for training the shared module. That is, after using x (x=1, 2, 3, 4, 5) new tasks to train the shared module, the parameters of the shared module are frozen. Since in the aforementioned experiments, the convolution kernel scale is (3*3) and the number of shared module channels is (64,32), and the convolution kernel scale is (3*3) and the number of shared module channels is (32,64), , the bearing fault diagnosis accuracy can reach more than 96%, and there is not much difference. Therefore, the convolution kernel scale is set to (3*3) in this section of the experiment. The two channel numbers and the learning times of the three shared modules are used to conduct combined experiments. The step size of the convolution in the experiment is 1. The experimental results are shown in the table 7.

Table 7 Diagnostic accuracy of shared modules with different learning times (Time)

It can be seen from Table 7 that when the learning number of the shared module is 1, the bearing fault diagnosis accuracy can reach the highest under the conditions of the two shared module channel numbers, so the optimal learning number of the shared module is 1.

The feature map generated by the number of shared module channels (64,32) is smaller than that of (32,64), and the complexity of the model is lower, resulting in excessive loss, gradient explosion, and training failure.

As the number of learning times (that is, the number of new tasks for training shared modules) increases, the accuracy of the model decreases. After the first task is trained, the first feature extractor has fitted the parameters of the shared module. When the second task continues to train the shared module, the parameters of the shared module are changed, causing the first feature extractor to fail. Better fit of shared module parameters. Therefore, the more training times, the accuracy will decrease. The shared module plays a role in extracting common basic features, regardless of the number of new tasks used for training.

It can be seen from the above ablation experiment results that when the convolution kernel scale of the shared module is set to (3*3), the number of channels is (64,32), and the number of learning times is set to 1, the accuracy of bearing fault diagnosis is the highest, reaching 96.7%.

Comparative analysis with mainstream network models

In order to compare the effects of SMDER and the classic incremental learning network model in the field of bearing fault diagnosis, SMDER and the incremental learning method based on model structure (DER), the incremental learning method based on playback (iCaRL), and the regularization-based incremental learning method were respectively Incremental learning method (LwF) was used to conduct fault diagnosis experiments on the CWRU bearing data set. The experimental results are shown in Table 8.

Table 8 Comparative analysis with mainstream incremental learning methods

It can be seen from Table 8 that although the training time of the SMDER network model proposed in this application is slightly longer than other methods, the bearing fault diagnosis accuracy that can be achieved is the highest. It can be seen that the basic features of time-frequency features extracted by the shared module play an important role in the structure of incremental learning.

Model reduction experiment

Since SMDER adds a new ResNet18 as a feature extractor for each new task, as the tasks continue to be added, the entire network model will occupy more and more memory and the training time will become longer and longer, so ResNet18 is repeated in the middle. The four two-layer convolutions that appeared were deleted and became ResNet10. The network model structure diagram of ResNet10 is shown in Figure 8.

The experimental results of using the reduced ResNet10 to replace ResNet18 as the feature extractor are shown in Table 9.

Table 9 Model reduction experimental results

As can be seen from Table 9, after the feature extractor is reduced from ResNet18 to ResNet10, the bearing fault diagnosis accuracy only drops by 0.08%, but the training time is reduced by 29 seconds.

The comparison of parameters between ResNet18 and ResNet10 is shown in Table 10.

Table 10 Comparison of parameters between ResNet18 and ResNet10

As can be seen from Table 10, the number of parameters of ResNet10 is reduced by more than half compared to ResNet18. Therefore, under incremental learning conditions, using ResNet10 as the feature extractor will reduce 6.3×10 ⁶ parameters when adding each task. Although using ResNet10 as a feature extractor will cause a slight decrease in fault diagnosis accuracy, it can reduce memory space usage and training time. Therefore, under the conditions of limited computing resources and training time, you can consider using the reduced feature extractor ResNet10.

Claims (5)

1. A rolling bearing fault diagnosis method based on SMDER, which is characterized by including the following steps: Pass the bearing vibration signal through continuous wavelet transform to obtain the time-frequency domain characteristics of the original signal, and then input the time-frequency domain characteristics into the shared module for processing; Input the features extracted by the shared module into the super feature extractor for feature extraction and then use the classifier for classification processing to achieve bearing fault diagnosis under incremental learning conditions; The shared module first inputs the wavelet time-frequency diagram into the first convolution layer of the shared module for convolution processing, then performs batch normalization BatchNorm, and uses the ReLU activation function to process, and then inputs the obtained feature vector into The second convolution layer, and finally batch normalization BatchNorm; after the shared module operation is completed, the obtained features will be input to the feature extractor of each task for processing according to the new tasks; The super feature extractor includes multiple feature extractors, each feature extractor corresponding to an incremental learning task, in which the old feature extractor is used to extract features of existing data. After the new class arrives, the existing feature extractor is frozen. parameters, and add a new feature extractor to extract the features of the new class. At this time, the combination of all feature extractors is a super feature extractor, and then the features extracted by all feature extractors are spliced together, which is a super feature extractor. The extracted features; the formula of the super feature extractor is as shown in Equation (1): z＝F _t (x)＝[F _t-1 (x), Y _t (x)] (1) In the formula, x is the feature output by the shared module and is also the input of the super feature extractor F _t ; the super feature extractor is spliced by the super feature extractor F _t-1 of the previous step and the new feature extractor Y _t of the current step. Composition; z is the feature extracted by the super feature extractor. 2. The rolling bearing fault diagnosis method based on SMDER as claimed in claim 1, characterized in that: the feature extractor is pruned using a differentiable channel-level mask method, and the channel-level mask and features are learned together. After completing the mask, the mask is binarized and the feature extractor is pruned to obtain the pruned network; the channel-level mask is shown in Equation (2): h′ _l ＝h _l ⊙σ(se _l ) (2) where h' _l is the mask feature after pruning; h _l is the mask feature before pruning; ⊙ is the channel-level multiplication; σ is the sigmoid gating function, whose value range is [0,1]; s is Scaling factor; el is a parameter of the mask, which can be learned during training. 3. The rolling bearing fault diagnosis method based on SMDER as claimed in claim 1, characterized in that: the loss function of the entire network is as shown in formula (3): In the formula, L is the network loss; is the training loss; λ _α is the hyperparameter of the auxiliary classifier, and the initial value is 0;/> is the auxiliary loss; λ _s is the hyperparameter that controls the size of the model; _LS is the sparse loss. 4. The rolling bearing fault diagnosis method based on SMDER as claimed in claim 1, characterized in that: the old feature parameters of the classifier _Ht are the parameters frozen in the step Ht _-1 used, and the parameters of the new features are randomly initialized. ; Input the feature z extracted by the super feature extractor to the classifier H _t for classification and prediction, as shown in Equation (4): y＝argmax(Softmax(H _t (z))) (4) In the formula, y is the final predicted classification result; argmax is the point set whose independent variable reaches the maximum value; Softmax is the activation function of the output layer, which converts the prediction result into a non-negative number; H _t is the classifier; z is the super feature Features extracted by the extractor. 5. A rolling bearing fault diagnosis system based on SMDER, which is characterized by including: The wavelet transformation module is used to transform the bearing vibration signal through continuous wavelet transformation to obtain the time-frequency domain characteristic map of the original signal; The sharing module is used to extract features from the input time-frequency domain feature map; the sharing module first inputs the wavelet time-frequency map into the first convolution layer of the sharing module for convolution processing, and then performs batch normalization BatchNorm , and use the ReLU activation function to process, and then input the obtained feature vector to the second convolution layer, and finally perform batch normalization BatchNorm; after the shared module operation is completed, the obtained features will be different according to the new tasks. The input is processed by the feature extractor of each task; Super feature extractor, including multiple feature extractors, each feature extractor corresponds to an incremental learning task, in which the old feature extractor is used to extract features of existing data. After the new class arrives, the parameters of the existing feature extractor are frozen. , and add a new feature extractor to extract the features of the new class. At this time, the combination of all feature extractors is a super feature extractor, and then the features extracted by all feature extractors are spliced together, which is the super feature extractor extraction. Characteristics; The classifier is used to classify the features extracted by the super feature extractor to realize bearing fault diagnosis under incremental learning conditions.