Flexible thin-wall bearing fault diagnosis method based on STFT and DCPCNN-SVM
By using STFT and DCPCNN-SVM in the fault diagnosis of flexible thin-wall bearings, the problem that diagnostic effects in the prior art are limited by expertise and experience, and high accuracy and robust fault diagnosis under small sample conditions are achieved.
Patent Information
- Application Number
- CN202411818075.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-10-17
- Filing Date
- 2024-12-11
- Publication Date
- 2025-06-24
AI Technical Summary
The diagnostic effect of prior art in the diagnosis of flexible thin-wall bearings is limited by the user's expertise and experience, and the application of deep learning models in this field has not yet reached its full potential, especially under small sample conditions.
The flexible thin-wall bearing fault diagnosis method based on STFT and DCPCNN-SVM is adopted, and the original signal is converted into a two-dimensional time-frequency image through short-time Fourier transform, and a two-channel parallel CNN model is used for fault diagnosis with a support vector machine.
Under small sample conditions, this method exhibits good diagnostic performance and robustness, can effectively extract fault characteristics with good distinction, and improves the accuracy and reliability of fault diagnosis of flexible thin-wall bearings.
Smart Images
Figure CN120197010A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flexible thin-wall bearing fault diagnosis, and particularly relates to a flexible thin-wall bearing fault diagnosis method based on STFT and DCPCNN-SVM. Background Art
[0002] Flexible Thin-Wall Bearing (FTWB) is a key transmission component in a robot harmonic reducer. Due to its elliptical structure, which is completely different from that of a conventional rolling bearing during operation, there is always a background impact component in its vibration signal caused by the alternation of the long and short axes of the bearing. Such impacts are a great interference to the fault impacts of the flexible thin-wall bearing. Therefore, studying the state monitoring and diagnosis methods of flexible thin-wall bearings is of great significance for the reliable and continuous operation of high-end mechanical equipment such as robots.
[0003] The key to diagnosing the faults of flexible thin-wall bearings lies in processing their vibration signals. The main methods involved include: Continuous Wavelet Transform (CWT), Ensemble Empirical Mode Decomposition (EEMD), Variational Mode Decomposition (VMD), Maximum Correlation Kurtosis Deconvolution (MCKD), Singular Value Decomposition (SVD), etc. Guo Yingying et al. proposed a fault feature extraction method combining Continuous Morlet Wavelet Transform (CMWT) and Hilbert envelope spectrum analysis, and successfully identified the inner and outer ring faults of flexible thin-wall bearings. Liu Xingjiao et al. used EEMD to decompose the vibration signal, then screened out the key IMF components according to the kurtosis principle to reconstruct the signal, and finally successfully extracted the outer ring fault features of flexible thin-wall bearings using parameter-optimized MCKD. Lu et al. used VMD to decompose the bearing vibration signal, reconstructed the signal based on the IMF determined by the maximum kurtosis principle, and then combined with Multi-Point Optimal Minimum Entropy Deconvolution Adjustment (MOMEDA) and envelope spectrum analysis to accurately extract the outer and inner ring fault characteristic frequencies of flexible thin-wall bearings. Zhao Xuezhi et al. analyzed the unique periodic impact characteristics of flexible thin-wall bearings and effectively separated their background impact characteristics according to the amplitude filtering characteristics of SVD.
[0004] Although the above methods have played an important role in the fault diagnosis of flexible thin-walled bearings, their diagnostic effects are often limited by the professional knowledge and experience of users, which to a certain extent restricts their universality and accuracy. In recent years, deep learning has been favored by researchers in many fields such as image processing, natural language processing, speech recognition, and autonomous driving due to its deep structure, which has extremely strong complex function relationship fitting performance and excellent adaptive feature learning ability. Deep learning models widely used in the field of mechanical fault diagnosis include: Deep Belief Network (DBN), Convolutional Neural Network (CNN), Stacked Denoising Auto-encoder (SDAE), Long Short-Term Memory (LSTM), etc. For example, Xu et al. extracted multi-dimensional features of the vibration signal of a rolling bearing through CWT and EEMD, and then optimized the structure and training parameters of DBN using the Sparrow Search Algorithm (SSA), significantly improving the accuracy and effectiveness of fault diagnosis. Xie Fengyun et al. converted the vibration signal into a two-dimensional time-frequency diagram through STFT, then used two-dimensional CNN for feature extraction, and finally input it into SVM for classification, effectively improving the accuracy and reliability of gearbox fault diagnosis. Chen et al. proposed an improved SDAE bearing fault diagnosis method, which significantly improved the diagnostic accuracy, anti-noise performance, and generalization ability of the model by introducing a structure adaptive adjustment strategy and a new comprehensive loss function. Gu et al. used Discrete Wavelet Transform (DWT) to extract multi-scale data from the vibration signals collected by different sensors and fuse them, and then input the fused data into the LSTM network for feature extraction to accurately identify 10 different fault types at the shaft end.
[0005] The above literature fully demonstrates the excellent performance of deep learning in the field of mechanical fault diagnosis such as rolling bearings, but there are only very few literature reports on the related research of flexible thin-walled bearings. For example, Shao Qipeng took the original time-domain vibration signal as the input and studied the effects of four deep learning methods (CNN, DBN, SAE, LSTM) on the fault diagnosis of flexible bearings, and found that only the diagnostic accuracy of the CNN model was close to 90%, and the others were relatively low. Guo Yingying proposed a weak fault diagnosis method based on sparse decomposition and the Sparse Auto-encoder (SAE) model, where sparse decomposition was used to extract fault impact features, and SAE was used as a fault classifier for fault classification. This method achieved good diagnostic results, but the feature extraction and classification recognition were divided into two independent stages, and the specific operation steps were relatively cumbersome, failing to fully utilize the advantages of deep learning models in automatic feature learning. Except for these two pieces of literature, no other literature has reported relevant content.
[0006] In summary, the research on intelligent diagnosis of flexible thin-walled faults by deep learning has just started. The existing research only stays at the stage of using the original time-domain signals and their statistical features as inputs, and has not involved frequency-domain features, time-frequency features, and other more complex image features. Summary of the Invention
[0007] The present invention aims to provide a flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM. This method is scientifically and reasonably designed and still has good diagnostic performance and robustness under small sample conditions.
[0008] The technical solution of the present invention is as follows:
[0009] The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM includes the following steps:
[0010] A. Collect multiple groups of original vibration signals of normal flexible thin-walled bearings and faulty flexible thin-walled bearings respectively. Use the sliding window processing method to perform overlapping sampling on the collected vibration signals, perform short-time Fourier transform (STFT) processing on all samples to generate STFT time-frequency images, and then uniformly process the picture size to obtain a training set.
[0011] B. Use a dual-channel parallel CNN, replace the Softmax classification layer of the traditional CNN with a support vector machine (SVM) to obtain a DCPCNN-SVM model, and perform supervised training on the model with the training set samples to obtain a trained DCPCNN-SVM model.
[0012] C. Collect the original vibration signal of the flexible thin-walled bearing to be tested, use the sliding window processing method to perform overlapping sampling on the collected vibration signals, perform short-time Fourier transform (STFT) processing on all samples to generate STFT time-frequency images, and then uniformly process the picture size to obtain a feature set.
[0013] D. Input the feature set into the trained DCPCNN-SVM model for testing to identify the fault type.
[0014] In step A, the faulty flexible thin-walled bearings include outer ring crack bearings, inner ring crack bearings, and outer and inner ring composite fault bearings.
[0015] In step A, when using the sliding window processing method for overlapping sampling, the step size is 1000, the number of sampling points is 2048, and 100 samples are collected for each type of fault.
[0016] In step A, the picture size is uniformly processed to 64x64x3.
[0017] In the said step B, the dual-channel parallel CNN consists of two CNN channels. The structures of the two CNN channels are the same. The input results enter the two channels for processing respectively. After the obtained results are concatenated by the Concatenate function, they are processed by the third fully connected layer and the support vector machine SVM in sequence to obtain the final fault diagnosis result.
[0018] The processing process in the said CNN channel is as follows:
[0019] The input result is processed by the first convolutional layer, batch normalization, Relu function, and max pooling layer in sequence, then passes through the Dropout layer to randomly discard some neurons, and then is processed by the second convolutional layer, batch normalization, Relu function, max pooling layer, first fully connected layer, and second fully connected layer in sequence to obtain the output result.
[0020] The calculation formula of the said short-time Fourier transform STFT is:
[0021]
[0022] In the formula, θ(t) represents the time-domain signal, and f(t - ω) represents the window function.
[0023] Compared with the prior art, the advantages of the present invention are:
[0024] The method of the present invention proposes a flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM, which can better cope with problems such as low fault diagnosis rate of deep learning in flexible thin-walled bearings and insufficient information content of input data. The short-time Fourier transform (STFT) is used to convert the original signal into a two-dimensional time-frequency image, so that the input data can contain more fault information. The analysis result of the t-SNE dimensionality reduction visualization of the last fully connected layer of the model shows that the proposed model can extract fault features with good discrimination.
[0025] By comparing the influence of different training sample numbers on the diagnosis effects of the DCPCNN-SVM and CNN-SVM models, it can be known that the method of the present invention still has good diagnosis performance and robustness under small sample conditions and has good application prospects. Description of the Drawings
[0026] Figure 1 It is the structure diagram of the neural network DCPCNN-SVM in Embodiment 1;
[0027] Figure 2 It is the test result diagram with 15 training samples of each type in Embodiment 2;
[0028] Figure 3 It is the test result diagram with 30 training samples of each type in Embodiment 2;
[0029] Figure 4 The number of training samples for each category in Example 2 is 50, which is the test result graph. Specific implementation manners
[0030] The present invention will be specifically described below in conjunction with the accompanying drawings and embodiments.
[0031] Example 1
[0032] A fault diagnosis method for flexible thin-walled bearings based on STFT and DCPCNN-SVM includes the following steps:
[0033] A. Collect multiple groups of original vibration signals of normal flexible thin-walled bearings and faulty flexible thin-walled bearings respectively. Use the sliding window processing method to perform overlapping sampling on the collected vibration signals, perform short-time Fourier transform (STFT) processing on all samples to generate STFT time-frequency images, and then uniformly process the picture size to 64x64x3 to obtain a training set;
[0034] The faulty flexible thin-walled bearings include outer ring crack bearings, inner ring crack bearings, and inner and outer ring compound fault bearings.
[0035] When using the sliding window processing method for overlapping sampling, the step size is 1000, the number of sampling points is 2048, and 100 samples are collected for each type of fault.
[0036] B. Use a dual-channel parallel CNN, replace the Softmax classification layer of the traditional CNN with a support vector machine (SVM) to obtain the DCPCNN-SVM model as shown in Figure 1 Perform supervised training on the model with the training set samples to obtain a trained DCPCNN-SVM model;
[0037] The dual-channel parallel CNN consists of two CNN channels. The structures of the two CNN channels are the same. The input results enter the two channels for processing respectively. After the obtained results are concatenated by the Concat function, they are sequentially processed by the third fully connected layer and the support vector machine (SVM) to obtain the final fault diagnosis result.
[0038] The processing process in the CNN channel is as follows:
[0039] The input results are sequentially processed by the first convolutional layer, batch normalization, Relu function, and max pooling layer, then pass through the Dropout layer to randomly discard some neurons, and then are sequentially processed by the second convolutional layer, batch normalization, Relu function, max pooling layer, first fully connected layer, and second fully connected layer to obtain the output result.
[0040] The calculation formula of the short-time Fourier transform (STFT) is:
[0041]
[0042] In the formula, θ(t) represents the time-domain signal, and f(t - ω) represents the window function.
[0043] C. Use the trained DCPCNN-SVM model to test the test set to identify the fault type.
[0044] Example 2
[0045] Comparative Experiment on Fault Diagnosis of Flexible Thin-Wall Bearings
[0046] 1 Test Conditions
[0047] The test data is sourced from the detection of an acceleration sensor. The outer ring fault of the measured flexible thin-wall bearing is simulated by a rectangular pit with a length and width of 1 mm processed by electrical discharge machining. The specific parameters are shown in Table 1.
[0048] Table 1 Basic Test Parameters
[0049]
[0050] 2 Fault Diagnosis of the Outer Ring of the Flexible Thin-Wall Bearing
[0051] Conduct diagnostic tests based on the method of Example 1:
[0052] Collect multiple groups of original vibration signals of normal flexible thin-wall bearings and faulty flexible thin-wall bearings. Use the sliding window processing method to perform overlapping sampling on the collected vibration signals, perform short-time Fourier transform (STFT) processing on all samples to generate STFT time-frequency images, and then uniformly process the image size to 64x64x3. Randomly select 15, 30, and 50 images from each type of sample to form the training set, that is, the training set sample sizes are composed of 60, 120, and 200 images respectively, and the corresponding test sets contain 340, 280, and 200 images respectively. Use the training set data with three sample sizes to train the model constructed by the method of Example 1 of the present invention and the single-branch CNN-SVM model respectively, and then use the corresponding test set to test the trained model. The results are as Figures 2 - 4 shown.
[0053] From Figure 2It can be seen that when the number of training samples is small, the DCPCNN-SVM model of the method in Embodiment 1 of the present invention has a higher accuracy rate in each test than that of the single-branch CNN-SVM. The accuracy rate of the former fluctuates between 83.53% and 92.35%, and its average accuracy rate is 88.08%; while the accuracy rate of the latter varies between 59.41% and 78.82%, and the average accuracy rate is 69.53%. Thus, it can be seen that in the case of small-sample training, the robustness and diagnostic performance of the model proposed in Embodiment 1 of the present invention are better than those of the single-branch CNN-SVM.
[0054] Figure 3 It shows that when the number of training samples is doubled, the diagnostic effects of both models have been greatly improved. Among them, the accuracy rate of the DCPCNN-SVM model of the method in Embodiment 1 varies between 96.07% and 99.29%, and the average accuracy rate reaches 97.5%. Compared with Figure 2 the previous situation, the accuracy rate has increased by 9.42 percentage points; while the accuracy rate of the single-branch CNN-SVM varies in the range of 66.07% - 84.29%, with a large fluctuation range, and the average accuracy rate is only 74.07%. Although it has also increased by about 4.54 percentage points relative to Figure 2 the previous situation, the diagnostic effect is still unsatisfactory. When the number of training samples increases to 50, at this time the ratio of training samples to test samples is 1:1, and the test results of the two models are as Figure 4 shown. It can be seen from the figure that the accuracy rate of the DCPCNN-SVM model of the method in Embodiment 1 fluctuates between 96% and 100%, and the average accuracy rate reaches 98.2%, and the diagnostic effect is still relatively ideal; while the accuracy rate of the single-branch CNN-SVM fluctuates between 81.5% and 98.5%. Although the average accuracy rate reaches 92.6%, its robustness is relatively poor.
[0055] It can be seen that by comparing the influence of different numbers of training samples on the diagnostic effects of the DCPCNN-SVM in Embodiment 1 of the present invention and the single-branch CNN-SVM model of the prior art, it can be known that the model in Embodiment 1 of the present invention still has good diagnostic performance and robustness under small-sample conditions; while the diagnostic effect of the single-branch CNN-SVM model is greatly affected by the number of samples. Especially under small-sample conditions, its diagnostic performance far fails to meet the requirements of actual engineering applications.
Claims
1. A flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM, characterized in that: The following steps are involved: A. Collect multiple groups of original vibration signals of normal flexible thin-walled bearings and faulty flexible thin-walled bearings respectively, use sliding window processing method to perform overlapping sampling on the collected vibration signals, perform short-time Fourier transform (STFT) processing on all samples to generate STFT time-frequency images, and then unify the image size to obtain the training set; B. Use dual-channel parallel CNN to replace the Softmax classification layer of the traditional CNN with a support vector machine SVM to obtain a DCPCNN-SVM model. Use the training set samples to perform supervised training on the model to obtain a trained DCPCNN-SVM model. C. Collect the original vibration signal of the flexible thin-walled bearing to be tested, use the sliding window processing method to perform overlapping sampling on the collected vibration signal, perform short-time Fourier transform (STFT) processing on all samples to generate STFT time-frequency images, and then unify the image size to obtain a feature set; D. Input the feature set into the trained DCPCNN-SVM model for testing to identify the fault type.
2. The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM as claimed in claim 1 is characterized in that: In the step A, the faulty flexible thin-walled bearing includes a bearing with an outer ring crack, a bearing with an inner ring crack, and a bearing with composite inner and outer ring faults.
3. The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM as claimed in claim 1 is characterized by: In step A, when the sliding window processing method is used for overlapping sampling, the step length is 1000, the number of sampling points is 2048, and 100 samples are collected for each fault.
4. The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM as claimed in claim 1 is characterized by: In step A, the image size is uniformly processed to 64x64x3.
5. The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM as claimed in claim 1 is characterized by: In the step B, the dual-channel parallel CNN is composed of two CNN channels, the two CNN channels have the same structure, and the input results are processed in the two channels respectively. The obtained results are concatenated by the Concatenate function, and then processed by the third fully connected layer and the support vector machine SVM in turn to obtain the final fault diagnosis result.
6. The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM as claimed in claim 5 is characterized by: The processing process in the CNN channel is as follows: The input result is processed by the first convolutional layer, batch normalization, Relu function, and maximum pooling layer in sequence, and then randomly discards some neurons through the Dropout layer. Then it is processed by the second convolutional layer, batch normalization, Relu function, maximum pooling layer, the first fully connected layer, and the second fully connected layer in sequence to obtain the output result.
7. The flexible thin-walled bearing fault diagnosis method based on STFT and DCPCNN-SVM as claimed in claim 1 is characterized by: The calculation formula of the short-time Fourier transform STFT is: Where θ(t) represents the time domain signal and f(t-ω) represents the window function.