Bearing fault diagnosis method based on improved lightweight deep convolutional neural network

Abstract

The present invention discloses a bearing fault diagnosis method based on an improved lightweight deep convolutional neural network, and belongs to the field of rolling bearing component fault monitoring in rotating mechanical equipment. First, the vibration signal collected by the sensor is divided into a plurality of data samples of equal length, and converted into image data, a fault label is marked for each type of sample, and a training data set and a test data set are divided; then the divided data set is used to train the constructed bearing fault diagnosis model to obtain a pre-trained model. The pre-trained model is pruned to eliminate redundant parameters in the model to obtain a lightweight model; finally, the pruned model is fine-tuned with the data set to restore the diagnostic performance of the model to obtain a lightweight model with high diagnostic performance. The bearing fault diagnosis model of the present invention has a small size, short training time, and fast response speed, and can accurately and more quickly identify bearing faults in actual industrial applications.

Description

Technical Field

[0001] The present invention relates to the field of rolling bearing component fault monitoring in rotating mechanical equipment, and in particular to a bearing fault diagnosis method based on an improved lightweight deep convolutional neural network. Background Art

[0002] In modern industrial systems, rotating machinery plays an important role in mechanical equipment and is widely used in various mechanical equipment. Rolling bearings are one of the core components of rotating machinery, and their health status is used to evaluate whether the mechanical system is working properly. Under the influence of various factors, bearing components will inevitably fail, leading to major accidents. In severe cases, property losses will be caused, and the life safety of mechanical operators will also be threatened. Therefore, it is very urgent to study fault diagnosis methods that can quickly and accurately monitor whether the machinery is running stably online.

[0003] In recent years, the powerful adaptive feature extraction capability of deep convolutional neural networks has made them widely used in the field of bearing fault diagnosis. The fault diagnosis method based on deep neural networks can adaptively extract fault features from raw data, effectively avoiding dependence on expert experience knowledge. In addition, this method integrates feature extraction and fault classification into a whole, which realizes adaptive feature extraction and fault classification. However, deep convolutional neural networks usually improve the feature extraction capability of the model by stacking multiple convolutional layers, and use fully connected layers with a large number of neurons to realize fault classification. This leads to a large number of parameters in the bearing fault diagnosis model based on deep convolutional neural networks, and requires high storage and computing capabilities for monitoring equipment. This greatly limits the application of such models in actual industrial scenarios.

[0004] Model compression methods can effectively reduce model complexity, thereby reducing the requirements for software and hardware of monitoring equipment. At present, technologies based on parameter pruning, parameter quantization, and lightweight model design have been used to solve the problem of large model parameters. For example, removing redundant parameters of the fully connected layer in the model can reduce the number of model parameters. Reducing the number of bits representing the parameters reduces the memory resources required to store the model. Using some lightweight convolution structures to replace the standard convolution in the traditional deep convolutional neural network can obtain a lightweight model and reduce the computational complexity of the model. However, the above methods rarely process the model from multiple aspects, and the number of model parameters is mainly concentrated in the fully connected layer. Even if some parameters are removed, it is difficult to avoid requiring large-capacity storage resources for the monitoring equipment. When more parameters are removed, the model diagnosis accuracy will be greatly reduced, affecting the model performance. The combination of multiple methods will prolong the training time of the model, thereby reducing the response speed of the model.

[0005] From the above analysis, it can be seen that it is difficult to obtain a lighter model by applying only one model compression method, and it is difficult to apply high-performance deep convolutional neural network models in actual industrial scenarios. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, in order to obtain a more lightweight deep convolutional neural network model, improve the response speed of the model, and reduce the requirements for storage and computing resources of the monitoring equipment. The present invention proposes a bearing fault diagnosis method based on an improved lightweight deep convolutional neural network. First, in the feature extraction module, a heterogeneous convolution structure is used to replace the traditional standard convolution structure, and a lightweight model based on heterogeneous convolution can be obtained through training; secondly, a pruning operation is introduced on the basis of the lightweight model based on heterogeneous convolution to process the redundant parameters that still exist in the heterogeneous convolution, eliminate channels with redundant filters, and further reduce the floating-point operations generated by the convolution operation; finally, in the classification module, the fault features output by the feature extraction structure are fused with the original features after Dropout processing, and the fused fault features are input into the global average pooling layer to achieve feature dimensionality reduction. The above structure is used to replace the fully connected layer in the traditional classification module, which greatly reduces the number of model parameters and speeds up the response speed of the model.

[0007] The present invention is achieved through the following solutions:

[0008] A bearing fault diagnosis method based on an improved lightweight deep convolutional neural network includes the following steps:

[0009] Step 1: The bearing vibration signal collected by the sensor is a one-dimensional signal. The collected signal is divided into multiple data samples of the same length at a certain interval, and converted into a two-dimensional grayscale image sample. A sample label is assigned to each sample, and the training data set and the test data set are divided;

[0010] Step 2: The lightweight deep convolutional neural network constructed is a bearing fault diagnosis model, including: a lightweight feature extraction module, a convolutional attention channel module, and an improved fault classification module;

[0011] Step 3: Initialize model parameters, including number of iterations, learning rate, weight parameters, bias, training batch-size, test batch-size, and optimizer;

[0012] Step 4: Train the constructed bearing fault diagnosis model through the training data set, update the model parameters through forward propagation and back propagation, train until the model converges, and save the model with the best performance;

[0013] Step 5: Input the test data set into the trained bearing fault diagnosis model, test the model, and verify the model's diagnostic accuracy for all fault categories.

[0014] The further improvement of this method is that the data is preprocessed in step 1, specifically in the following way:

[0015] The collected vibration signal is divided into multiple one-dimensional samples with a sample length of n, that is, each sample contains n data points; for a one-dimensional data sample, the process of converting it into a two-dimensional grayscale image can be described by the following formula:

[0016]

[0017] Where L(k) is the original one-dimensional data intercepted at a certain interval, with a length of n, k = 1, 2, ..., n, Represents the pixel value of the converted image at point (i, j), where n is a positive integer greater than 1. The data sample can contain multiple fault categories. By adjusting the radial force generated by the hydraulic loading system and the speed generated by the AC motor in the data acquisition device, data samples under different working conditions can be collected to obtain data of multiple fault categories. Multiple categories are arbitrarily selected to obtain the total number of samples m, 70% of which are used as training data and 30% as test data.

[0018] The further improvement of this method is that the step 2 is specifically:

[0019] (1) The lightweight feature extraction module consists of three pruned heterogeneous convolutional layers and one standard convolutional layer to extract the complex features of fault data.

[0020] (2) A convolutional attention module is introduced to optimize the fault features output by the feature extraction module to improve model performance.

[0021] (3) The improved fault classification module consists of a Dropout and global average pooling. Before the Dropout is input into the global average pooling, it is fused with the original feature data to improve the model classification performance.

[0022] The further improvement of this method is that the step three is specifically:

[0023] The training parameters of the model are set as follows: the number of training iterations is 50, the initial learning rate is 0.1, the learning rate is decayed tenfold at iterations 30 and 40 respectively, the initial weight parameters and bias are initialized to random numbers, the batch-size of the training dataset is 64, the batch-size of the test dataset is 32, and the optimizer is Adam.

[0024] The further improvement of this method is that the step 4 is specifically:

[0025] The training process of the bearing fault diagnosis model is as follows: first, the initial model is trained using the training data set, and the model performance is tested using the test data set. When the expected number of iterations is reached and the model converges, the optimal model after convergence is saved. Then, the redundant channels in the trained model are removed through the set evaluation criteria to obtain a lightweight model, and the model performance is restored by fine-tuning again.

[0026] The further improvement of this method is that the channel evaluation standard of the bearing fault diagnosis model is the l1-norm size of the filter in each output channel, and its calculation formula is as follows:

[0027]

[0028] Where n i is the number of input channels, K j is the size of the jth convolution kernel in the input channel.

[0029] A further improvement of this method is that the model of the lightweight feature extraction module is:

[0030]

[0031] Among them, M w ,M h ,C i are the size of the input feature map and the number of input channels, W i and b i is the weight and bias of the i-th convolutional layer, Y i The output pwc and gwc of the i-th convolutional layer are pointwise convolution and groupwise convolution respectively.

[0032] A further improvement of this method is that the model of the convolutional attention channel module is:

[0033] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0034] =σ(W1(W0(F avg ))+W1(W0(F max )))

[0035]

[0036] Among them, M c (F),M s(F) represents the channel attention output and spatial attention output, respectively. 7×7 Represents the size of the convolution kernel, σ represents the sigmoid function, W0 and W1 represent the weight parameters in the multi-layer perceptron (MLP), AvgPool is the average pooling operation, and MaxPool is the maximum pooling operation.

[0037] The further improvement of this method is that the model of the improved fault classification module is:

[0038] o=Softmax(AvgPool(Dropout(F)+F))

[0039] Among them, o is the fault classification output result vector, F is the convolutional attention module output feature matrix, and AvgPool is the average pooling operation.

[0040] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0041] The present invention discloses a bearing fault diagnosis method based on an improved lightweight deep convolutional neural network. The method constructs a bearing fault diagnosis model with very few parameters by constructing a lightweight feature extraction module and a classification module. By converting the original one-dimensional signal into a two-dimensional grayscale image, the complex fault features in the fault data can be better extracted. Compared with the existing technology, the present invention can obtain a lightweight model with high diagnostic performance, reduce the software and hardware requirements of the model deployment on the monitoring equipment, speed up the response speed and training speed of the model, and realize the rapid diagnosis of bearing faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 It is a flow chart of the implementation method of the present invention;

[0044] Figure 2 It is a schematic diagram of the arrangement of the bearing fault data acquisition device of the present invention;

[0045] Figure 3 It is the two-dimensional grayscale image corresponding to ten one-dimensional fault signals;

[0046] Figure 4 It is the overall model structure diagram of the present invention;

[0047] Figure 5 The improved classification module diagram of the present invention;

[0048] Figure 6 This is a confusion matrix diagram of the lightweight model of the present invention for ten fault diagnosis results. DETAILED DESCRIPTION

[0049] The present invention will be described in detail below with reference to the accompanying drawings:

[0050] like Figure 1 , a bearing fault diagnosis method based on an improved lightweight deep convolutional neural network, comprising:

[0051] (1) The bearing vibration signal collected by the sensor is a one-dimensional signal. The collected signal is divided into multiple data samples of the same length at a certain interval and converted into a two-dimensional grayscale image sample, where Figure 3 Grayscale images after transformation of ten fault signals. Assign sample labels to each sample and divide the training data set into test data sets;

[0052] In this embodiment, the bearing fault data set published by Xi'an Jiaotong University is directly used to describe the specific implementation steps of the present invention. The data set is collected by the Mechanical Equipment Health Monitoring Joint Laboratory established by the School of Mechanical Engineering of Xi'an Jiaotong University and Zhejiang Changxing Shengyang Technology Co., Ltd.

[0053] The data collected are vibration signals, with a sampling frequency of 25.6kHz and a sampling interval of 1 minute. By adjusting the radial force generated by the hydraulic loading system and the speed generated by the AC motor, data of ten different fault types were obtained.

[0054] The bearing vibration data collected above are marked with different labels, as shown in Table 1:

[0055]

[0056]

[0057] In the above table, the number of each type of fault is 300, the total number of samples is 3000, the number of samples in the training data set is 2100, and the number of samples in the test data set is 900.

[0058] (2) Construct a lightweight deep convolutional neural network bearing fault diagnosis model including: lightweight feature extraction module, convolutional attention channel module and improved fault classification module.

[0059] (2-1) The lightweight feature extraction structure consists of three pruned heterogeneous convolutional layers and one standard convolutional layer, which is used to extract the complex features of fault data.

[0060] (2-2) A convolutional attention module is introduced to optimize the fault features output by the feature extraction structure to improve model performance.

[0061] (2-3) The improved fault classification module consists of a Dropout and a global average pooling. Before the Dropout is input into the global average pooling, it is fused with the original feature data to improve the model classification performance. The structure is as follows: Figure 5 shown.

[0062] (3) Using the training data set and the test data set to train the model, including: obtaining a pre-trained model, model pruning, and model fine-tuning.

[0063] (3-1) Set the hyperparameters for model training: the number of training iterations is 50, the initial learning rate is 0.1, the learning rate is decayed by ten times at iterations 30 and 40 respectively, the initial weight parameters and bias are initialized to random numbers, the batch-size of the training dataset is 64, the batch-size of the test dataset is 32, and the optimizer is Adam.

[0064] (3-2) Use the training data set to train the initial model, and use the test data set to test the model performance. When the expected number of iterations is reached and the model converges, save the optimal model after convergence.

[0065] (3-3) Redundant channels in the trained model are removed through the set evaluation criteria to obtain a lightweight model, and the model performance is restored by fine-tuning again. The model channel evaluation criterion is the l1-norm size of the filter in each output channel, and its calculation formula is as follows:

[0066]

[0067] where n i is the number of input channels, K j is the size of the jth convolution kernel in the input channel.

[0068] (4) The trained lightweight and high-performance diagnostic model is deployed to actual industrial monitoring equipment to achieve real-time rapid bearing fault diagnosis.

[0069] Experimental example:

[0070] Figure 6 The results of ten fault categories are shown in Figure 1. It can be seen that the recognition accuracy is 95% or above. The parameter volume of the model obtained by the present invention is only 0.02M, the floating point operation volume is only 1.64M, and the model training time is only 226.57 seconds. It can be seen that the obtained model is very small, the model training time is short, and the response speed is fast. In the real monitoring scenario, the bearing fault category can be diagnosed faster.

[0071] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A bearing fault diagnosis method based on an improved lightweight deep convolutional neural network, characterized in that: The method comprises the following steps: Step 1: The bearing vibration signal collected by the sensor is a one-dimensional signal. The collected signal is divided into multiple data samples of the same length at a certain interval, and converted into a two-dimensional grayscale image sample. A sample label is assigned to each sample, and the training data set and the test data set are divided; Step 2: The lightweight deep convolutional neural network constructed is a bearing fault diagnosis model, including: a lightweight feature extraction module, a convolutional attention channel module, and an improved fault classification module; In the lightweight feature extraction module, a heterogeneous convolution structure is used to replace the traditional standard convolution structure, and a lightweight model based on heterogeneous convolution is obtained through training; on the basis of the lightweight model based on heterogeneous convolution, a pruning operation is introduced to process the redundant parameters that still exist in the heterogeneous convolution, remove channels with redundant filters, and further reduce the floating-point operations generated by the convolution operation; in the improved fault classification module, the fault features output by the feature extraction structure are fused with the original features after Dropout processing, and the fused fault features are input into the global average pooling layer to achieve feature dimensionality reduction; specifically: The lightweight feature extraction module consists of three pruned heterogeneous convolutional layers and one standard convolutional layer. The convolutional attention module is used to optimize the fault features output by the feature extraction module to improve model performance. The improved fault classification module consists of a Dropout and global average pooling, and before the Dropout is input into the global average pooling, it is fused with the original feature data. The model of the lightweight feature extraction module is: HetConv=pwc+gwc Among them, M w ,M h ,C i are the size of the input feature map and the number of input channels, W i and b i For the i The weights and biases of the convolutional layers, Y i The output pwc and gwc of the i-th convolutional layer are point-by-point convolution and group convolution respectively; The model of the convolutional attention channel module is: M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) =σ(W1(W0(F avg ))+W1(W0(F max ))) Among them, M c (F),M s (F) represents the channel attention output and spatial attention output, respectively. 7×7 represents the size of the convolution kernel, σ represents the sigmoid function, W0 and W1 represent the weight parameters in the multi-layer perceptron (MLP), AvgPool represents the average pooling operation, and MaxPool represents the maximum pooling operation; The model of the improved fault classification module is: o=Softmax(AvgPool(Dropout(F)+F)) Among them, o is the fault classification output result vector, F is the convolutional attention module output feature matrix, and AvgPool is the average pooling operation; Step 3: Initialize model parameters, including number of iterations, learning rate, weight parameters, bias, training batch-size, test batch-size, and optimizer; Step 4: Train the constructed bearing fault diagnosis model through the training data set, update the model parameters through forward propagation and back propagation, train until the model converges, and save the model with the best performance; The training of the bearing fault diagnosis model includes: training of the diagnosis model based on heterogeneous convolution and fine-tuning of the pruned model; first, the initial model is trained using the training data set, and the model performance is tested using the test data set. When the expected number of iterations is reached and the model converges, the optimal model after convergence is saved; then, the redundant channels in the trained model are removed through the set evaluation criteria to obtain a lightweight model, and the model performance is restored by fine-tuning again; The channel evaluation standard of the bearing fault diagnosis model is the l1-norm size of the filter in each output channel, and its calculation formula is as follows: Among them, n i is the number of input channels, K j is the size of the jth convolution kernel in the input channel; Step 5: Input the test data set into the trained bearing fault diagnosis model, test the model, and verify the model's diagnostic accuracy for all fault categories.

2. The bearing fault diagnosis method based on improved lightweight deep convolutional neural network according to claim 1 is characterized in that: In step 1, the length of the generated one-dimensional sample is n, that is, each sample contains n data points; for a one-dimensional data sample, the process of converting it into a two-dimensional grayscale image can be described by the following formula: Where L(k) is the original one-dimensional data intercepted at a certain interval, with a length of n, k = 1, 2, ..., n, Represents the pixel value of the transformed image at point (i, j).

3. The bearing fault diagnosis method based on improved lightweight deep convolutional neural network according to claim 1 is characterized in that: In the step three, the number of training iterations is 50, the initial learning rate is 0.1, the learning rate is decayed tenfold when the number of iterations is 30 and 40, respectively, the initial weight parameters and bias are initialized to random numbers, the batch-size of the training data set is 64, the batch-size of the test data set is 32, and the optimizer is selected as Adam.

Images