Rolling bearing fault classification diagnosis method and device
By using a fault classification and diagnosis model with multi-scale residual modules and channel attention mechanism, the problem of low accuracy in rolling bearing fault classification under noisy backgrounds is solved, and high-accuracy fault classification is achieved in strong noise environment.
Patent Information
- Application Number
- CN202411123277.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2026-03-03
Smart Images
Figure CN121595205A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rolling bearing fault technology, specifically relating to a method and apparatus for classifying and diagnosing rolling bearing faults. Background Technology
[0002] Rotating systems are an important component of rotating machinery, and bearings are key components of rotating systems. They are prone to failure and difficult to inspect in the early stages. Once a failure occurs, the rotating machinery will lose some or all of its functions, causing economic losses and even personal injury or death. Therefore, it is very important to detect and diagnose the operating condition of bearings.
[0003] The operating signals of rolling bearings are nonlinear and easily affected by background noise. Traditional signal processing methods struggle to identify faults from complex vibration signals. Machine learning-based intelligent diagnostic methods are an effective approach, directly extracting feature information from the acquired vibration signals to identify faults, thus avoiding reliance on expert knowledge. However, these machine learning models often have shallow network structures, making it difficult to extract representative fault features amidst background noise, resulting in low diagnostic accuracy. Summary of the Invention
[0004] The purpose of this invention is to provide a method and apparatus for classifying and diagnosing rolling bearing faults, in order to solve the problem of low accuracy in the classification and diagnosis of rolling bearing faults under noisy conditions in the prior art.
[0005] To address the aforementioned technical problems, this invention provides a method for classifying and diagnosing rolling bearing faults. The method includes: acquiring vibration data of the rolling bearing and inputting it into a trained fault classification and diagnosis model to obtain a rolling bearing fault classification and diagnosis result; wherein the fault classification and diagnosis model includes a feature extraction module and a classification module; the feature extraction module includes a multi-scale residual module, which is used to extract features of different scales from the vibration data using at least two convolutional units of different scales, and to fuse the extracted features of different scales; wherein each convolutional unit includes a residual block or at least two sequentially connected residual blocks, and each residual block uses a convolutional layer for feature extraction; the classification module is used to classify the fault based on the result output by the feature extraction module.
[0006] Furthermore, the feature extraction module also includes a wide convolutional layer, which is used to suppress high-frequency noise in the input vibration data and input the processed vibration data into the multi-scale residual module so that the multi-scale residual module can extract features from the vibration data after suppressing high-frequency noise.
[0007] Furthermore, the fault classification and diagnosis model also includes a feature filtering module. The feature filtering module filters the output of the multi-scale residual module through a channel attention mechanism and inputs the filtered features into the classification module so that the classification module can classify faults based on the filtered features.
[0008] Furthermore, the feature fusion is a feature splicing process.
[0009] Furthermore, the training data used to train the fault classification and diagnosis model includes vibration data with Gaussian white noise added at different signal-to-noise ratios.
[0010] Furthermore, the loss function used to train the fault classification and diagnosis model is the cross-entropy loss function.
[0011] Furthermore, the number of convolutional units is three.
[0012] Furthermore, the training and testing data used to train and test the fault classification and diagnosis model include vibration data processed from the original vibration data obtained through overlap sampling technology.
[0013] To address the aforementioned technical problems, the present invention also provides a rolling bearing fault classification and diagnosis device, comprising a processor that executes a computer program to implement the steps of the rolling bearing fault classification and diagnosis method described above.
[0014] Its beneficial effects are as follows: This invention is an improved invention. Existing rolling bearing fault diagnosis methods have simple structures and are difficult to extract effective fault features under background noise, resulting in inaccurate diagnosis results. This invention obtains fault classification and diagnosis results by inputting the acquired vibration data of the rolling bearing into a fault classification and diagnosis model. The fault classification and diagnosis model includes a feature extraction module and a classification module. The feature extraction module includes a multi-scale residual module, which is used to extract features from the vibration data through at least two convolutional units of different scales, and fuse the extracted features of different scales. Each convolutional unit... It includes one residual block or at least two sequentially connected residual blocks, each residual block using a convolutional layer for feature extraction; here, residual blocks are used for feature extraction, and the residual blocks extract input features based on residual learning and skip connections. Residual learning enables the network to learn the correlation between shallow and deep features of vibration data more quickly, even if there are huge differences between the input and output, which can improve the convergence speed and generalization ability of the network; skip connections can directly connect the input layer to the output layer, so that the residual can be directly passed to subsequent layers, preventing the vanishing gradient caused by the increase of network layers, and accelerating the training and convergence of the model. In noisy operating conditions, accurate fault classification requires minimizing the impact of noise during feature extraction. The skip connections of residual blocks help the network better retain input information, thereby reducing the interference of noise on feature extraction. Therefore, this classification module can accurately classify and diagnose faults based on the fault features after fusing shallow and deep features of the input data. This invention uses residual blocks for feature extraction, which can effectively improve the robustness of the model and accurately classify faults in the absence of noise, especially under noisy operating conditions. This improves the accuracy of the entire fault classification and diagnosis model in the noisy background of rolling bearing fault classification and diagnosis. Attached Figure Description
[0015] Figure 1 This is a diagnostic flowchart of a rolling bearing fault classification and diagnosis method according to an embodiment of the present invention;
[0016] Figure 2 This is a schematic diagram of the multi-scale residual module according to an embodiment of the present invention;
[0017] Figure 3 This is a schematic diagram of the channel attention mechanism module according to an embodiment of the present invention;
[0018] Figure 4 This is a schematic diagram of the overlap sampling technique according to an embodiment of the present invention;
[0019] Figure 5 This is a diagram showing the confusion matrix diagnosis results of rolling bearing data under noise-free conditions according to an embodiment of the present invention.
[0020] Figure 6 This is a graph showing the t-SNE diagnostic results of rolling bearing data under noise-free conditions according to an embodiment of the present invention;
[0021] Figure 7 This is a diagram showing the confusion matrix diagnosis results of rolling bearing data under strong noise according to an embodiment of the present invention;
[0022] Figure 8 This is a t-SNE diagnostic result diagram of rolling bearing data under strong noise according to an embodiment of the present invention;
[0023] Figure 9 This is a diagram showing the confusion matrix diagnosis results of existing CNN models for rolling bearing data under strong noise.
[0024] Figure 10 This is a diagram showing the diagnostic results of the existing MSCNN model for the confusion matrix of rolling bearing data under strong noise. Detailed Implementation
[0025] This invention obtains fault classification and diagnosis results by inputting the acquired vibration data of rolling bearings into a fault classification and diagnosis model. The model includes a feature extraction module and a classification module. The feature extraction module includes a multi-scale residual module, which extracts features from the vibration data using at least two convolutional units of different scales and fuses the extracted features at different scales. Each convolutional unit includes one residual block or at least two sequentially connected residual blocks, and each residual block uses a convolutional layer for feature extraction. The classification module performs fault classification based on the output of the feature extraction module. This invention's fault classification and diagnosis model can learn the correlation between shallow and deep features more quickly during feature extraction, improving the accuracy of rolling bearing fault classification and diagnosis results in noisy environments.
[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0027] Method Implementation Examples:
[0028] This embodiment provides a method for classifying and diagnosing rolling bearing faults, such as... Figure 1 As shown, the specific steps include:
[0029] Step 1: Construct a fault classification and diagnosis model. This model includes a feature extraction module, a feature selection module, and a classification module. The feature extraction module comprises wide convolutional layers and a multi-scale residual module. The structure, function, and interrelationships of each module are described below.
[0030] (1) Wide convolutional layer:
[0031] The wide convolutional layer is a 64*1*16 wide convolution. The input vibration data first passes through the wide convolutional layer to expand the network's receptive field. By increasing the kernel size, the wide convolutional layer can better suppress high-frequency noise compared to small convolutional kernels. The wide convolutional layer is defined as follows:
[0032]
[0033] in, The output feature information of the l-th layer, f(·) is the activation function, and M j Represents the input feature mapping. The input information of the l-th layer is the output feature information of the (l-1)-th layer; For convolution kernel weights, For bias, I and j are used to represent positional information in the convolution operation.
[0034] (2) Multi-scale residual module:
[0035] like Figure 2 As shown, the multi-scale residual module extracts features at different scales from the vibration data (i.e., features obtained after wide-kernel convolutional layers) processed by wide convolutional layers using three convolutional units of different scales, and then fuses the extracted features at different scales (feature fusion in this embodiment is feature concatenation). The three convolutional units are respectively... Figure 1 The convolutional layers consist of 1×3 convolutions / residuals, 1×5 convolutions / residuals, and 1×7 convolutions / residuals, with kernel sizes of 3*1*8, 5*1*8, and 7*1*8 respectively. In this embodiment, there are three convolutional units. In other implementations, the number and size of the convolutional units can be adjusted according to actual needs. Each convolutional unit includes three sequentially connected residual blocks. Each residual block uses a convolutional layer for feature extraction to better utilize feature information at different scales, helping the multi-scale residual module learn the correlation between shallow and deep features more quickly, thus improving the convergence speed and generalization ability of the multi-scale residual module. In this embodiment, there are three residual blocks. In other implementations, the number can be adjusted according to actual needs. The residual blocks are specifically represented as follows:
[0036] x2=F{x1,(W)}+x1 (2)
[0037] Where x1 represents the input feature, x2 represents the output of input x1 after passing through the residual network, F{x1,(W)} represents the learned residual mapping function, and W represents the weight parameters of the residual block.
[0038] This embodiment utilizes residual blocks for feature extraction to improve model performance, especially in noisy conditions, enabling accurate fault classification. Residual learning in residual blocks refers to the model learning a representation of the residuals (differences) between the input and output within each residual block. Even with significant differences between the input and output, the network can more easily learn these residuals, thus more easily learning task-relevant features. This residual learning mechanism helps the network learn features useful for fault classification more effectively.
[0039] Furthermore, the convolutional layers in the residual blocks all use skip connections. Skip connections directly connect the input and output layers, allowing residuals to be directly propagated to subsequent layers. This helps solve the problems of vanishing and exploding gradients, and also accelerates model training convergence. In noisy environments, accurate fault classification requires minimizing the impact of noise during feature extraction. The skip connections in the residual blocks help the network better preserve input information, thus reducing noise interference in feature extraction. Therefore, using residual blocks for feature extraction effectively improves the model's robustness, enabling the network to accurately classify faults even in noisy conditions.
[0040] (3) Feature Filtering Module:
[0041] The feature filtering module is used to filter features output by the multi-scale residual module. In this embodiment, the feature filtering module uses a channel attention mechanism (i.e., Figure 1 The mid-channel attention mechanism feature enhancement module performs feature filtering on the output of the multi-scale residual module. This can be selected as another implementation method based on actual needs. Through the channel attention mechanism, the entire model can dynamically adjust the weights of features according to the importance of different scales, thereby better capturing the global and local information of the feature sequence. For example... Figure 3 As shown, the channel attention mechanism module is implemented as follows:
[0042] ① The feature sequence of each channel is compressed by using a global average pooling layer. The calculation formula is as follows:
[0043]
[0044] Where Avgpool represents global flat pooling operation, x i Let W represent the feature sequence, W represent the dimensional information of the input matrix, and K represent the feature sequence. i This represents the characteristics after pooling operations.
[0045] ② Embed global time information into K, and the specific calculation method is as follows:
[0046] K'=β(F”(α(F'(K)))) (4)
[0047] Where K' represents the feature vector calculated by the attention mechanism, and K represents the features before channel attention, i.e., Ki i α represents the activation function; F' and F” represent two fully connected layers, F' performs dimensionality reduction, and F” restores its dimensionality and encodes its channels; β is the sigmoid function, which assigns a weight to each channel.
[0048] ③ Multiply the weight generated for each channel by each feature of the previous layer. The specific calculation formula is as follows:
[0049] Y = F scale (K′·X)=K′×X (5)
[0050] Among them, F scale Y represents the scaling operation, Y represents the output of the channel attention mechanism module, and X represents the input before the attention mechanism.
[0051] (4) Classification module:
[0052] The classification module is used to classify faults based on the features input from the feature filtering module.
[0053] Step 2: Obtain sample data and preprocess the obtained sample data, including vibration data of rolling bearings.
[0054] To improve data diversity and generalization ability, overlapping samples are extracted from the original vibration data. By setting the offset and sample length, different fault vibration signals are overlapped using an equally spaced offset sliding method, thereby increasing the number of sample data. In this embodiment, the sample length in the overlap sampling technique is 2048, and the offset is 10. A schematic diagram of the overlap sampling technique is shown below. Figure 4 As shown. Alternatively, the offset and sample length can be set according to actual needs.
[0055] To improve the accuracy of the fault classification and diagnosis model constructed in step one, Gaussian white noise with different signal-to-noise ratios can be added to the vibration data used to train the model to simulate a noisy environment. The noise level is represented by the signal-to-noise ratio (SNR), as shown below:
[0056]
[0057] Among them, X signal and X noise This represents the original signal power and the noise signal power.
[0058] Step 3: Divide the sample data processed in Step 2 into a training set and a test set, and use the training set and test set to train and test the fault classification and diagnosis model built in Step 1 respectively.
[0059] In this method, the original dataset is divided into two parts: a training set and a test set. The training set is used to build the model, and the test set is used to evaluate the model's accuracy and generalization ability. Data splitting can avoid overfitting and improve the model's stability and generalization ability.
[0060] To improve the accuracy of the fault classification and diagnosis model, it is optimized using the backpropagation method based on error calculation. Furthermore, the loss function used to train the fault classification and diagnosis model is the cross-entropy loss function, which can help the model to deeply extract effective features of vibration data under noise interference.
[0061] This embodiment inputs the original signals from the test set into the model and observes the model's output. The accuracy and performance of the model are evaluated by visualizing the confusion matrix and t-SNE. The confusion matrix is used to determine the number of samples correctly classified, and the t-SNE is used to determine whether the data features are clustered and classified.
[0062] In a confusion matrix, accuracy refers to the proportion of correctly predicted samples out of the total number of samples. The horizontal axis represents the predicted fault type, and the vertical axis represents the fault label. t-SNE visualization is a dimensionality reduction algorithm that maps high-dimensional data to a two- or three-dimensional space for visualization. By mapping it to a low-dimensional space using t-SNE, the distribution of samples of different categories in the low-dimensional space can be observed, providing a visual assessment of the model's ability to distinguish between different categories and thus understanding the model's performance on the dataset. The x and y axes in T-SNE represent the position of data points in the low-dimensional space, determined by the similarity between high-dimensional data points. Closer data points are also closer in the low-dimensional space. Different colors represent different fault type features; data points of the same color are grouped together, indicating correct classification.
[0063] Step 4: Use the fault classification and diagnosis model trained in Step 3 to perform fault classification and diagnosis, and obtain the classification results.
[0064] The following are the diagnostic results of the rolling bearing fault classification and diagnosis method of the present invention applied to specific examples, such as... Figure 5 The figure shown is a diagram of the confusion matrix diagnosis results under noise-free conditions according to an embodiment of the present invention. It can be seen that, except for a few misjudgments in two states, the predictions of other fault types are all correct, with multiple tests achieving an accuracy of over 99.5%. Figure 6The image shows the t-SNE diagnostic results of an embodiment of the present invention under noise-free conditions. The t-SNE visualization technology clearly displays the classification results, demonstrating that after processing the original data using the invented method, all data features are clearly classified and clustered. This effectively illustrates that the present invention has excellent fault identification capabilities.
[0065] like Figure 7 , Figure 8 The figures shown are the confusion matrix diagnostic results and t-SNE diagnostic results under strong noise interference according to an embodiment of the present invention. The experimental results are also visualized using the confusion matrix and t-SNE. Figure 8 This displays the classification results at SNR=2dB. From Figure 7 The confusion matrix diagnostic results show that only a few cases were misjudged, and the accuracy rate of multiple tests reached over 98%. Furthermore, it can be seen that various faults are accurately separated in the t-SNE visualization.
[0066] like Figure 9 The image shown is a confusion matrix diagnostic result diagram for rolling bearing fault classification using a CNN model under strong noise interference in existing technologies; from Figure 9 As can be seen, when there are 200 fault types "8", only 155 are correctly identified as fault type "8", while the remaining 45 are incorrectly identified as fault type "4", resulting in an accuracy rate of only 77.5%; the accuracy rate for identifying fault type "9" is also only 77.5%. Figure 10 The image shown is a confusion matrix diagnostic result of the MSCNN model used in the prior art for rolling bearing fault classification under strong noise interference; from Figure 10 As can be seen, when there are 200 fault types "4", only 163 were correctly identified as fault type "4", while the remaining 37 were incorrectly identified as fault type "7", with an accuracy rate of only 81.5%.
[0067] Comparing the classification results of various models, it can be seen that the fault classification and diagnosis method of the present invention can not only achieve an extremely high accuracy rate in fault classification in the absence of noise, but also accurately identify the fault type and classify it in a strong noise environment, thus meeting the practical application requirements.
[0068] In summary, the multi-scale residual attention mechanism network model proposed in this invention addresses the problem of poor diagnosis due to noise interference. Compared to other signal processing techniques, this method achieves end-to-end intelligent diagnosis and exhibits higher diagnostic accuracy compared to shallow machine learning models. Based on this model, rolling bearing fault types can be effectively diagnosed even in strong noise environments, providing a new technique for preventing rolling bearing failures.
[0069] Device Example:
[0070] An embodiment of the present invention provides a rolling bearing fault classification and diagnosis device, comprising a memory, a processor, and an internal bus. The processor and the memory communicate and interact with each other via the internal bus. The memory includes at least one software function module stored in the memory. The processor executes various functional applications and data processing by running the software program and module stored in the memory, thereby implementing the rolling bearing fault classification and diagnosis method described in the method embodiment of the present invention. The principle, implementation process, and achievable effects of this method have been fully described in the method embodiment and will not be repeated here.
Claims
1. A method for classifying and diagnosing rolling bearing faults, characterized in that, The method includes: acquiring vibration data of rolling bearings and inputting it into a trained fault classification and diagnosis model to obtain rolling bearing fault classification and diagnosis results; wherein, the fault classification and diagnosis model includes a feature extraction module and a classification module; the feature extraction module includes a multi-scale residual module, which is used to extract features of different scales from the vibration data through at least two convolutional units of different scales, and to fuse the extracted features of different scales; wherein each convolutional unit includes a residual block or at least two residual blocks connected in sequence, and each residual block uses a convolutional layer for feature extraction; the classification module is used to perform fault classification based on the results output by the feature extraction module.
2. The rolling bearing fault classification and diagnosis method according to claim 1, characterized in that, The feature extraction module further includes a wide convolutional layer, which is used to suppress high-frequency noise in the input vibration data and input the processed vibration data into the multi-scale residual module so that the multi-scale residual module can extract features from the vibration data after suppressing high-frequency noise.
3. The rolling bearing fault classification and diagnosis method according to claim 1, characterized in that, The fault classification and diagnosis model also includes a feature filtering module. The feature filtering module filters the output of the multi-scale residual module through a channel attention mechanism and inputs the filtered features into the classification module so that the classification module can classify faults based on the filtered features.
4. The rolling bearing fault classification and diagnosis method according to claim 1, characterized in that, The feature fusion is a feature splicing process.
5. The rolling bearing fault classification and diagnosis method according to claim 1, characterized in that, The training data used to train the fault classification and diagnosis model includes vibration data with Gaussian white noise added at different signal-to-noise ratios.
6. The rolling bearing fault classification and diagnosis method according to claim 1, characterized in that, The loss function used to train the fault classification and diagnosis model is the cross-entropy loss function.
7. The rolling bearing fault classification and diagnosis method according to claim 1, characterized in that, The number of convolutional units is 3.
8. The rolling bearing fault classification and diagnosis method according to any one of claims 1 to 7, characterized in that, The training and testing data used to train and test the fault classification and diagnosis model include vibration data processed from the original vibration data obtained through overlap sampling technology.
9. A rolling bearing fault classification and diagnosis device, comprising a processor, characterized in that, The processor executes a computer program to implement the steps of the rolling bearing fault classification and diagnosis method as described in any one of claims 1 to 8.