Bearing fault diagnosis method based on dual-channel multi-scale attention feature fusion

Through the dual-channel multi-scale attention feature fusion method, combining vibration signals and reconstruction images, depth feature information is extracted, and the problem of insufficient feature extraction and training time in the prior art is solved, and efficient fault diagnosis under variable working conditions is achieved .

CN118918424APending Publication Date: 2024-11-08NANJING UNIV OF POSTS & TELECOMM

Patent Information

Application Number
CN202410938988.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing deep learning bearing fault diagnosis model has insufficient feature extraction and training time, especially in variable operating conditions, which makes it difficult to accurately identify fault signals, resulting in unsatisfactory diagnostic accuracy and speed.

Method used

The dual-channel multi-scale attention feature fusion method is adopted, combining the original vibration signal and the reconstruction image, and the depth feature information is extracted through the multi-scale attention network, and the weight is adaptively allocated through the feature fusion module to reduce training time and improve testing accuracy.

Benefits of technology

It significantly improves the accuracy and speed of bearing fault diagnosis, especially the fault recognition ability under variable operating conditions, reduces training time, and enhances the convergence speed of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918424A_ABST
    Figure CN118918424A_ABST
Patent Text Reader

Abstract

The invention discloses a bearing fault diagnosis algorithm based on dual-channel multi-scale attention feature fusion. The bearing fault diagnosis algorithm mainly solves the problem that fault diagnosis and cross-working-condition fault recognition are low in precision. According to the method, two channels are adopted to perform feature extraction on an original vibration signal and a reconstructed image, features of different semantics are extracted through a multi-scale convolutional layer, different weights are distributed to the channels and the space through an attention mechanism, useful information is captured, and interference feature information is suppressed; and fusing features through an iAFF adaptive feature fusion module, and identifying 10 fault categories. Experiments show that compared with a traditional fault model, the method has the advantages that the recognition precision is greatly improved, particularly, the method has high precision in cross-working-condition fault recognition, and the method has a great promotion effect on fault diagnosis of the bearing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of rolling bearing fault diagnosis, and in particular to a detection algorithm for bearing fault diagnosis based on deep learning technology. Background Art

[0002] With the increasing sophistication and complexity of rotating machinery, more stringent requirements are placed on the operating reliability of the equipment. Due to the complex operating conditions and harsh environment of rotating machinery, its components are extremely vulnerable to damage. In actual working conditions, any minor fault may cause the failure of components or the entire machine, and may even lead to devastating damage and catastrophic accidents. Therefore, it is crucial to effectively, quickly and accurately evaluate and identify the fault state of rotating machinery to improve its safety and reliability. As a key component of rotating machinery, the working reliability of rolling bearings will directly affect the accuracy and life of the entire equipment. Therefore, the bearing fault data collected by a variety of sensors in all directions provides a new learning method for the field of industrial diagnosis. At present, the fault diagnosis of motor bearings mainly uses acceleration sensors to collect vibration signals at motor bearings, and then uses intelligent algorithms to diagnose the signals. In recent years, deep learning methods have developed rapidly, and bearing fault diagnosis based on deep learning has gradually received widespread attention. Its powerful feature extraction capability overcomes the time-consuming and labor-intensive disadvantages of traditional diagnosis. The application of this technology can effectively improve the accuracy and speed of fault diagnosis.

[0003] Common deep learning models for bearing fault diagnosis include convolutional neural networks (CNN) and residual networks (ResNet). Typical models include multi-scale convolutional neural networks (MSCNN) and ResNet18, which are fast to train but have low accuracy and are prone to losing important feature information. In terms of feature fusion algorithms, concat and add fusion methods are more commonly used. These methods increase the training time to a certain extent and cannot make the model converge quickly. Therefore, designing an effective feature fusion method can effectively reduce the training time. In the field of fault diagnosis, variable operating condition migration is still a key issue. Since the fault signals under different operating conditions are relatively complex, it is difficult to extract key feature information, and it is often impossible to accurately identify the vibration signals of variable operating condition bearing faults. Realizing variable operating condition bearing migration fault diagnosis can greatly reduce the training time and reduce the collection of fault samples, which is of great significance to fault diagnosis research. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a bearing fault diagnosis method based on dual-channel multi-scale attention feature fusion, which combines the dual-channel method composed of the original vibration signal and the reconstructed image to retain the fault feature information to the greatest extent, extracts the deep feature information through the multi-scale attention network, and finally fuses the different semantic features through the feature fusion module, and adaptively assigns weights to the global and local feature information, so as to quickly and accurately extract multi-scale feature information, reduce training time, improve test accuracy, and have a good recognition ability for bearing faults under variable working conditions.

[0005] In order to achieve the above-mentioned purpose of the present invention, the technical solution adopted by the present invention is:

[0006] The bearing fault diagnosis method based on dual-channel multi-scale attention feature fusion includes the following steps:

[0007] Step 1: Overlap sampling of the CWRU bearing dataset driver data using the sliding window method, labeling and standardizing the sampled data, and dividing the training set, validation set, and test set into a ratio of 6:2:2.

[0008] Step 1 specifically includes:

[0009] Step 1.1 saves the vibration data of the drive end of 10 bearing working conditions, namely outer ring fault, inner ring fault, ball six o'clock fault and normal working condition. The sampling sizes of the three fault conditions are cracks of 0.007, 0.012, and 0.021 inches, respectively. Therefore, 10 different state categories are combined, including 9 fault categories and 1 normal category.

[0010] Step 1.2 Collect 400 samples for each fault category and normal state, for a total of 4000 samples, and divide the training set, validation set, and test set into 2400 training samples, 800 validation samples, and 800 test samples in a ratio of 6:2:2. Set the step size to 150 and the sequence length to 1024 to overlap the data samples using the sliding window method. The validation set samples are divided from the end position of the training set, and the test set samples are divided from the end position of the validation set, ensuring that there is no overlap between the training set, validation set, and test set.

[0011] Step 1.3: Label the divided training set, validation set, and test set and save them into a list, and standardize all data for subsequent neural network training.

[0012] Step 2 reconstructs the divided samples into two-dimensional images through the Gram angle difference field image generation technology, sets the pixel size of the generated two-dimensional image to 224×224, and saves the corresponding labels for subsequent neural network training.

[0013] Step 3 builds a dual-channel multi-scale attention feature fusion neural network. Both the signal channel and the image channel use multi-scale convolutional neural networks with different convolution kernel sizes for feature extraction. The feature output of each channel is further extracted through attention mechanism modules such as SE and CBAM. The dual channels are fused through iAFF, and finally 10 working state classification results are given.

[0014] Step 3 is as follows:

[0015] Step 3.1 uses a dual-channel input model, and each channel uses a convolutional neural network of different scales for feature extraction. The one-dimensional signal channel uses SE and ECA attention mechanism modules to further focus on the key features extracted by the multi-scale convolutional neural network. The specific operations are as follows: Figure 1 The 2D image channel uses the CBAM channel spatial attention mechanism to further focus on the key features extracted by the multi-scale convolutional neural network and generate the corresponding feature map. The specific operation is shown in the attached figure. Figure 2 shown.

[0016] Step 3.1 specifically includes:

[0017] Step 3.1.1: Use a multi-scale convolutional neural network to extract features from the input signal, and assign weights to different channels through the SE attention mechanism module. Then use the multi-scale convolutional neural network again to further extract features, and use the ECA attention mechanism to improve feature representation capabilities and reduce computational costs, and use the concat fusion method to fuse feature outputs from different branches.

[0018] In step 3.1.2, a multi-scale convolutional neural network is used to extract features from the input image information. The feature output is passed through the CBAM attention mechanism module to allocate attention weights to its channel space, thereby extracting deep image feature maps of different scales.

[0019] Step 3.2: Adaptively fuse the features of different scales extracted from the two channels in steps 3.1.1 and 3.1.2 through the iAFF fusion module, and set the channel reduction to 8. The specific structure is shown in the attached figure. Figure 3 As shown in the figure. The fused features are transformed nonlinearly through the nonlinear activation function sigmoid of the fully connected layer, and the input features are further extracted and combined to capture the complex patterns and relationships in the input data. Finally, the output of a fully connected layer is converted into a probability distribution through the softmax function to represent the probability of each category.

[0020] Step 3.3 uses the cross entropy loss function as the classification loss function. The calculation formula is as follows;

[0021]

[0022] Where N is the number of samples in the batch, C is the total number of categories, and y ni Represents the indicator function of the actual category of the nth sample, Represents the probability of the i-th category predicted by the model for the n-th sample.

[0023] Step 4 uses the designed network structure to train the training set, verifies the model parameters through the validation set and saves the optimal parameters, and finally tests on the test set to obtain the final test accuracy and diagnosis results.

[0024] Step 4 specifically includes:

[0025] Step 4.1 Use the stochastic gradient descent Adam optimizer to optimize the designed network structure, set the batchsize to 10, the learning rate to 0.0001, the weight decay to 1e-4, and the epoch to 15 rounds (batchsize: the number of samples obtained in each training, epoch: use all the samples in the training set to train once), and train the network until the network converges. Verify the model parameters through the validation set and save the optimal training parameters.

[0026] Step 4.2 uses the model weights obtained in step 4.1 to perform result tests on the test set to obtain the final test accuracy and diagnosis results.

[0027] Compared with other existing diagnostic techniques, the present invention has the following significant advantages:

[0028] 1. The Gram angle difference field image reconstruction technology is used to fully explore the correlation and similarity of the original vibration signal, retain the important features of the original signal to a great extent, capture the key information of the signal, and make full use of the superiority of convolutional neural network for image feature extraction.

[0029] 2. The attention mechanism allows the model to selectively focus on important parts while ignoring irrelevant or minor parts when processing input data. This selective attention can improve the accuracy and efficiency of the model. In natural language processing, the attention mechanism can effectively handle long-distance dependency problems. The iAFF adaptive feature fusion neural network is used to fuse the different scale features extracted by dual channels, and the training convergence speed is improved compared with the commonly used feature fusion methods such as concat and add.

[0030] 3. The network structure of the present invention can effectively identify the fault status of bearings across different working conditions, greatly reducing the training time required for different working conditions and solving the problem of fault diagnosis under variable working conditions to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1Image channel feature extraction module

[0032] Figure 2 Signal channel feature extraction module

[0033] Figure 3 Feature fusion module

[0034] Figure 4 Dual channel diagnostic model

[0035] Figure 5 Confusion matrix for the test set DETAILED DESCRIPTION

[0036] The present invention will be further explained in detail below in conjunction with the accompanying drawings and implementation methods. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0037] In order to enable those skilled in the art to better understand the implementation of the present invention, the present invention will use the Python simulation language to conduct a bearing fault diagnosis experiment based on dual-channel multi-scale attention feature fusion to verify its feasibility and reliability.

[0038] The specific implementation of the present invention is as follows:

[0039] Step 1: Overlap sampling of the CWRU bearing dataset is performed using the sliding window method, and the training set, validation set, and test set are divided into a ratio of 6:2:2.

[0040] Step 1 specifically includes:

[0041] Step 1.1 Obtain the drive-end vibration acceleration dataset in the CWRU dataset file and save the drive-end vibration data of 10 bearing working conditions, namely outer ring fault, inner ring fault, ball fault, and normal working condition. The three fault conditions are manifested as cracks with sizes of 0.007, 0.012, and 0.021 inches, so there are 10 different working condition categories.

[0042] Step 1.2 Collect 400 samples for each fault category, for a total of 4000 samples, and divide the training set, validation set, and test set into a ratio of 6:2:2, i.e., 2400 training samples, 800 validation samples, and 800 test samples.

[0043] Step 1.3 sets the sliding window method with a step size of 150 and a sequence length of 1024 to collect data samples. The validation set samples are divided from the end of the training set, and the test set is divided from the end of the validation set to ensure that there is no overlap between the training set, validation set, and test set to ensure test fairness.

[0044] Step 1.4 labels and standardizes the divided data samples. The dimension of the divided time series signal is set to (10, 1, 1024) as the input x1 of the signal channel.

[0045] Step 2 reconstructs the divided data samples into a two-dimensional image through the Gram angle difference field image generation technology, sets the generated image to 224×224 pixel size and RGB three channels, and sets the data dimension to (10, 224, 224, 3) as the input x of the image channel.

[0046] Step 3: Design the network structure and build a dual-channel multi-scale attention feature fusion neural network. Both the signal channel and the image channel use multi-scale convolutional neural networks with different convolution kernel sizes for feature extraction. The features of each channel are further extracted through attention mechanism modules such as SE, ECA, and CBAM. The dual channels are fused through iAFF, where iAFF is an adaptive feature fusion module. Finally, 10 fault classification results are given. The specific process is shown in the attached figure. Figure 4 shown.

[0047] Step 3 specifically includes:

[0048] Step 3.1 Input image x into the image channel, first use the convolution layer with a convolution kernel size of 16×16 for preliminary feature extraction, expand the number of channels from 3 to 16, and then compress the image size to 112×112 through the pooling layer to reduce the amount of training calculations and accelerate model convergence. Then, multi-scale convolution is performed through 4 branches to extract multi-scale information. The size of each convolution kernel is 16×16. After passing through the pooling layer, the feature maps of the four branches are spliced ​​to obtain a feature output of size (10, 64, 56, 56). The feature output is assigned different weights through the CBAM channel space attention mechanism, and finally through the convolution pooling layer, a feature output of size (10, 128, 7, 7) is obtained.

[0049] Step 3.2 Input signal x1 enters the signal channel, first use the convolution layer with a convolution kernel size of 64×64 to initially extract feature information, and then use the pooling layer to obtain the signal feature output with a size of (10, 32, 240). Then multi-scale convolution is performed through three branches, and the convolution kernel sizes are 5×5, 7×7, and 9×9 respectively. The features of the three branches are concatenated, and different weights are assigned to the channels through the SE attention mechanism. The feature output is again multi-scale convolution to obtain more important feature information, and finally the feature output of the signal channel is obtained through the ECA attention mechanism.

[0050] Step 3.3: The signal features and image features of the same size are fused through the iAFF adaptive feature fusion module to fuse the multi-scale features with different semantics. Finally, a dropout layer is added to the fully connected layer, and the inactivation rate p is set to 0.4.

[0051] Step 3.4 uses the cross entropy loss function as the classification loss function. Record the loss value and training accuracy of each training. The loss value calculation formula can be expressed as:

[0052]

[0053] Where N is the number of samples in the batch, C is the total number of categories, and y ni Represents the indicator function of the actual category of the nth sample, Represents the probability of the i-th category predicted by the model for the n-th sample.

[0054] Step 4 uses the network of the algorithm of the present invention to train the bearing training set, tests the accuracy through the verification set and saves the optimal model parameters, and uses the optimal model to detect the test set to obtain the final diagnosis result.

[0055] Step 4 specifically includes:

[0056] Step 4.1 Use Adam optimizer to optimize network parameters, set batchsize to 10, learning rate to 0.0001, weight decay to 1e-4, and epoch to 15. (batchsize: the number of samples obtained in each training, epoch: training once with all samples in the training set) Shuffle the order of the training set, ensure that the input signal, image and label correspond one to one, and train the network until the model converges and stabilizes.

[0057] Step 4.2 is verified by the validation set. If the accuracy of the validation set is greater than the accuracy of the training set, the parameter weight file of the training model is saved.

[0058] Step 4.3 Use the model weights obtained in step 4.2 to test the test set and record the test accuracy. The test confusion matrix is ​​as shown in the attached figure. Figure 5 shown.

[0059] Under the same conditions, multiple fault diagnosis methods are compared. MSCNN: Multi-scale Convolutional Neural Network; Res-Net: Residual Neural Network.

[0060] Table 1: Comparative test of different methods

[0061] MSCNN Res-Net 1-D CNN Algorithm of the present invention Recognition accuracy 94 95 90 98.5

[0062] Under the same conditions, the model is subjected to variable working condition migration experiments to verify whether different fault types can be accurately identified under variable working conditions. Therefore, four different load working condition data of 0hp, 1hp, and 2hp are selected under the CWRU data set for cross-load fault diagnosis.

[0063] AB means training and verification on dataset A and testing on dataset B. A total of 4 migration experiments were conducted, as shown in the following table:

[0064] A: 0hp B: 1hp C: 2hp

[0065] Table 2: Cross-operating condition diagnostic experiments

[0066] AB AC BC CA Average precision Accuracy 95.2 97 100 95.6 96.95

[0067] As can be seen from the above table, the present invention achieves good fault diagnosis accuracy in each group of cross-load fault diagnosis, and the average fault diagnosis accuracy is 96.95%. It can be seen that the present invention still has strong generalization ability under different working conditions.

[0068] The bearing fault diagnosis method based on dual-channel multi-scale attention feature fusion proposed in the present invention can improve the training speed and diagnosis accuracy, and has good results in variable working condition bearing fault diagnosis experiments.

[0069] The above is only a specific implementation of the present invention, but the protection scope of the present invention is not limited to this. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, can replace or change the technical solution and inventive concept of the present invention by equivalent means or all other embodiments obtained without making any creative work, all belong to the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on dual-channel multi-scale attention feature fusion, the specific steps are as follows: Step 1: Overlap sampling of the CWRU bearing dataset is performed using the sliding window method, and the training set, validation set, and test set are divided into a ratio of 6:2:

2. Step 2 reconstructs the divided samples into two-dimensional images through the Gram angular field image generation technology, and sets the pixel size of the generated image to 224*224 for subsequent neural network training. Step 3 builds a dual-channel multi-scale attention feature fusion neural network. Both the signal channel and the image channel use multi-scale convolutional neural networks with different convolution kernel sizes for feature extraction. The features of each channel are further extracted through attention mechanism modules such as SE and CBAM. Then, the dual channels are fused through the iAFF adaptive feature fusion module, and finally 10 fault classification results are given. Step 4 uses the network of the algorithm of the present invention to train the bearing training set, saves the optimal model parameters and tests the accuracy through the validation set to obtain the final diagnosis result.

2. According to claim 1, a bearing fault diagnosis method based on dual-channel multi-scale attention feature fusion, It is characterized in that In step 1, the sliding window overlapping sampling method is used to obtain sufficient data samples to avoid the problem of sample quantity imbalance.

3. According to the fault diagnosis method based on dual-channel multi-scale attention feature fusion as described in claim 1, It is characterized in that In step 2, the Gram angle difference field image generation technology is used to fully capture the correlation and similarity of the one-dimensional original vibration signal, and to make full use of the superior image processing capability of the convolutional neural network. Step 2 includes the following steps: Step 2.1 preprocesses the data and loads the data from the specified path. The parameters specify that the length of each sample is 1024, the number of samples of each type is 400, a total of 4000 fault samples are collected, the data is standardized, and 60% is specified as the training set, 20% as the validation set, and 20% as the test set. Step 2.2 uses the Gram Angle Difference Field (GADF) method to transform the test set data and convert the time series data into images. Step 2.3 generates and saves an image of the converted data in a fixed size. One image is generated for each sample and saved to a specified path.

4. According to the fault diagnosis method based on dual-channel multi-scale feature fusion as claimed in claim 1, It is characterized in that The step 3 comprises the following steps: Step 3.1 uses a dual-channel input model, and each channel uses a convolutional neural network of different scales for feature extraction. The one-dimensional signal channel uses SE and ECA attention mechanisms to further focus on the key features extracted by the multi-scale convolutional neural network. The two-dimensional image channel uses the CBAM channel spatial attention mechanism to further focus on the key features extracted by the multi-scale convolutional neural network. Step 3.2 uses the iAFF fusion module to perform adaptive feature fusion on the features of different scales extracted from the two channels. The fused features are nonlinearly transformed through the nonlinear activation function Sigmoid of the fully connected layer, and the input features are further extracted and combined to capture the complex patterns and relationships in the input data. Finally, the output of a fully connected layer is converted into a probability distribution through the softmax function to represent the probability of each category. Step 3.3 uses the cross entropy loss function as the classification loss function, and its calculation formula is as follows: Where N is the number of samples in the batch, C is the total number of categories, and y ni Represents the indicator function of the actual category of the nth sample, Represents the probability of the i-th category predicted by the model for the n-th sample.

5. According to the fault diagnosis method based on dual-channel multi-scale feature fusion as claimed in claim 1, It is characterized in that The step 3.1 comprises the following steps: Step 3.1.1 For the one-dimensional signal channel, first use a 64*64 convolution kernel to perform the first feature extraction on the initial signal, expand the number of channels to 32, and then use a multi-scale feature extraction network with convolution kernels of sizes 5*5, 7*7, and 9*9 to expand the number of channels to 64. Step 3.1.2 introduces the SE attention mechanism to adaptively assign different weights to each channel, emphasizing important feature channels and suppressing unimportant feature channels. The multi-scale feature extraction network in step 3.1.1 is used again to expand the number of channels to 128. The improved channel attention mechanism of ECA is used to achieve attention allocation between channels through one-dimensional convolution, avoiding the fully connected layer and simplifying the channel interaction process. Step 3.1.3 For the two-dimensional image channel, four branches are used to perform a 16*16 convolution kernel to extract multi-scale feature maps. The CBAM channel spatial attention mechanism is used to further focus on important feature maps, and then two convolution pooling layers are used to compress the size of the channel and feature map to further reduce the amount of calculation.

6. According to the fault diagnosis method based on dual-channel multi-scale attention feature fusion as claimed in claim 1, It is characterized in that The step 4 mainly includes the following steps: Step 4.1 Use the stochastic gradient descent Adam optimizer to optimize the designed network structure, set the batchsize to 10, the learning rate to 0.0001, the weight decay to 1e-4, the epoch to 15 rounds (batchsize: the number of samples obtained in each training, epoch: use all the samples in the training set to train once), and save the best training model and weight file. Step 4.2 uses the model weights obtained in step 4.1 to test the results on the test set.

Citation Information

Patent Citations

  • Multi-scale feature fusion gearbox fault diagnosis method based on self-attention mechanism

    CN116010900A

  • Double-channel convolution and BiGRU small-amount sample fault diagnosis method fusing attention mechanism

    CN117290697A

  • Planetary gearbox fault diagnosis method based on CNN and Transform combination

    CN117347046A

Cited By

  • Bearing fault diagnosis method, device and equipment and storage medium

    CN119290399A

  • A bearing fault diagnosis method, device, equipment and storage medium

    CN119290399B

  • Fault diagnosis method and system based on multi-scale dual feature enhancement

    CN119577535A

  • Fault diagnosis method and system based on multi-scale dual feature reinforcement

    CN119577535B

  • Bearing fault diagnosis method based on wavelet transform and mixed attention mechanism

    CN120277463A