Bearing fault diagnosis method and system based on semi-supervised contrast learning
By adopting a semi-supervised contrast learning method in bearing fault diagnosis, combining binary classifiers and closed-set classifiers, using BYOL network and RGB color vibration images, the problems of insufficient marking samples and unknown fault types are solved, and bearing fault diagnosis is improved The performance and accuracy are suitable for real industrial scenarios.
Patent Information
- Application Number
- CN202411431318.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-10-14
AI Technical Summary
In real industrial scenarios, the existing bearing fault diagnosis method based on deep learning is insufficient in the number of marking samples and the neglect of unknown fault types, resulting in a degradation in model performance and unable to effectively deal with a variety of fault types.
Using a semi-supervised contrast learning method, combining binary classifiers and closed-set classifiers, pseudo-labels are generated as open-set classification goals, and a feature extractor is pre-trained by BYOL self-supervised contrast learning network, and it is performed in a small amount of labeled data. Supervised pre-training, combined with RGB color vibration images for bearing fault diagnosis.
The performance of bearing fault diagnosis is improved when there are few labeled samples, and it can effectively deal with unknown fault types, improve the generalization ability and diagnostic accuracy of the model, and is suitable for real industrial scenarios.
Smart Images

Figure CN119086068A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a vibration signal processing method, a semi-supervised learning method and a comparative learning method, and belongs to the field of bearing fault diagnosis, and specifically relates to a bearing fault diagnosis method and system based on semi-supervised comparative learning. Background Art
[0002] Rolling bearings are the most widely used general mechanical components in rotating machinery. They play a very important role in rotating machinery. Whether their operating status is normal often directly affects the operation of the entire system. The rolling relationship between the main components of the bearing makes it the most vulnerable component in the machinery. Bearing failure may lead to high maintenance costs, serious economic losses and safety hazards. Therefore, intelligent diagnosis of bearing faults, as an important part of fault prediction and health management, is of great significance to ensure the safe and stable operation of rotating machinery and is a worthwhile research topic.
[0003] The data sets involved in intelligent diagnosis of bearing faults are mostly time series signals collected by sensors, and involve a lot of domain knowledge. Bearing fault diagnosis methods based on machine learning rely on a lot of expertise and complex feature engineering. In contrast, deep learning can directly extract and learn features from raw input data without introducing manually designed features or domain expertise. Therefore, many researchers have tried to apply deep learning methods to intelligent diagnosis of bearing faults and have achieved great success. However, fault diagnosis methods based on deep learning usually require a large amount of labeled data, but in real industrial scenarios, only a small part of the large amount of data obtained is often labeled. The problem of insufficient number of labeled samples has become an obstacle in this field. Therefore, semi-supervised learning methods have gradually become a research hotspot in bearing fault diagnosis.
[0004] In recent years, researchers have proposed some semi-supervised learning methods for bearing fault diagnosis. Although the proposed methods have improved the performance of bearing fault diagnosis to a certain extent when labeled samples are scarce, they ignore the possibility of unknown fault types, namely outliers (OOD). In actual industrial scenarios, the types of fault diagnosis are diverse, and there are many emergencies that will generate new fault types. These undefined fault data will become outliers in the training data, thereby reducing the performance of the model and making the method fail to achieve the expected results. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention proposes a bearing fault diagnosis method and system based on semi-supervised contrastive learning. An open set classifier is introduced under the traditional semi-supervised bearing fault diagnosis framework, and a binary classifier and a closed set classifier are combined to generate pseudo labels as open set classification targets.
[0006] To achieve the above object, the present invention provides the following solutions:
[0007] A bearing fault diagnosis method based on semi-supervised contrastive learning comprises the following steps:
[0008] Collecting original CWRU bearing vibration signals, and dividing the original CWRU bearing vibration signals into a training set and a test set, wherein the training set includes: a known class and an unknown class;
[0009] Preprocessing the original CWRU bearing vibration signal to obtain an RGB color vibration image;
[0010] Based on the RGB color vibration image, a bearing fault diagnosis model is established;
[0011] Fine-tuning the bearing fault diagnosis model;
[0012] Fault diagnosis is performed on the bearing to be tested based on the fine-tuned bearing fault diagnosis model.
[0013] Preferably, the method of preprocessing the original CWRU bearing vibration signal to obtain an RGB color vibration image comprises:
[0014] The original CWRU bearing vibration signal data is divided into data with a length of 1024 and covered, to obtain n vibration signal samples with a length of 1024;
[0015] Time Domain Conversion, STFT and CWT are performed on each vibration signal sample to obtain three two-dimensional grayscale images;
[0016] The three obtained 2D grayscale images are merged to generate an RGB color vibration image.
[0017] Preferably, the method for establishing a bearing fault diagnosis model based on the RGB color vibration image comprises: two stages;
[0018] Phase 1: Use BYOL self-supervised contrastive learning network to pre-train an effective bearing data feature extractor;
[0019] The second stage: Use a small amount of labeled training set data of known classes to pre-train a closed-set classifier in a supervised manner.
[0020] Preferably, the method of pre-training an effective bearing data feature extractor using a BYOL self-supervised contrastive learning network comprises:
[0021] The RGB color vibration image labels of the known class training set and test set are input into the BYOL network, and the enhanced views v and v′ are obtained after two data augmentations;
[0022] Input v into the online network, and pass through the ResNet50 network as a feature extractor and the MLP network as a projector in turn to obtain the representation y and projection z corresponding to the vibration image after data enhancement processing;
[0023] Similarly, v′ is input into the target network to obtain representations y′ and z′;
[0024] Input the projection z into the MLP network as a predictor to get the predicted value q θ (z).
[0025] Preferably, the method of using a small amount of labeled training set data of known classes to supervisely pre-train a closed set classifier includes:
[0026] Freeze the parameters of the pre-trained bearing data feature extractor during training;
[0027] The processed CWRU data set is input into the closed set classifier to obtain the probability of each extracted fault type, and the highest probability is taken as the classification result of the closed set classifier.
[0028] Preferably, the method for fine-tuning the bearing fault diagnosis model includes:
[0029] In the final stage of training, the pre-trained feature extractor and closed set classifier are used, the parameters of the feature extractor are frozen, and the closed set classifier and open set classifier are trained with a preset amount of labeled known class data and a preset amount of unlabeled known class and unknown class data.
[0030] The present invention also provides a bearing fault diagnosis system based on semi-supervised contrast learning, comprising: an acquisition module, a preprocessing module, a training module, a fine-tuning module and a diagnosis module;
[0031] The acquisition module is used to acquire the original CWRU bearing vibration signal and divide the original CWRU bearing vibration signal into a training set and a test set, wherein the training set includes: a known class and an unknown class;
[0032] The preprocessing module is used to preprocess the original CWRU bearing vibration signal to obtain an RGB color vibration image;
[0033] The training module is used to establish a bearing fault diagnosis model based on the RGB color vibration image;
[0034] The fine-tuning module is used to fine-tune the bearing fault diagnosis model;
[0035] The diagnosis module is used to perform fault diagnosis on the bearing to be tested based on the fine-tuned bearing fault diagnosis model.
[0036] Compared with the prior art, the present invention has the following beneficial effects:
[0037] In actual industrial scenarios, due to the low frequency of faults, the collected vibration signal fault samples are far less than the healthy state samples. In addition, the cost of manual labeling is high, which makes the labeled fault data more scarce. At the same time, a large amount of unlabeled fault data may contain unknown fault types, which can affect the diagnostic performance. Considering these problems, the present invention proposes a semi-supervised learning method (OSCL) based on contrastive learning for bearing fault diagnosis, which uses a small amount of known class labeled data and a large amount of unlabeled data containing known classes and unknown classes for joint training, which is more suitable for intelligent diagnosis of bearing faults in real industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. 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 labor.
[0039] Figure 1 A flow chart of a bearing fault diagnosis method based on semi-supervised contrastive learning according to an embodiment of the present invention;
[0040] Figure 2 A schematic diagram of a data preprocessing process according to an embodiment of the present invention;
[0041] Figure 3 A schematic diagram of a feature extractor training process according to an embodiment of the present invention;
[0042] Figure 4 Schematic diagram of a closed set classifier training process according to an embodiment of the present invention;
[0043] Figure 5 It is a schematic diagram of a flow chart of specific implementation steps of an embodiment of the present invention. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0046] Embodiment 1
[0047] like Figure 1 As shown, the present invention provides a bearing fault diagnosis method based on semi-supervised contrastive learning, comprising the following steps:
[0048] The original Case Western Reserve University (CWRU) bearing vibration signal is collected, and the original CWRU bearing vibration signal is divided into a training set and a test set, wherein the training set includes: a known class and an unknown class;
[0049] Preprocess the original CWRU bearing vibration signal to obtain RGB color vibration image;
[0050] Based on RGB color vibration images, a bearing fault diagnosis model is established;
[0051] Fine-tune the bearing fault diagnosis model;
[0052] Fault diagnosis of the bearing to be tested is performed based on the fine-tuned bearing fault diagnosis model.
[0053] The specific implementation process is as follows: Figure 5 As shown, data preparation: the original CWRU vibration signal is divided into training set and test set, and the training set data is processed separately, and aligned for CWT, STFT and Time Domain Conversion processing to obtain three 2D grayscale images, and the three grayscale images are merged to obtain an RGB color vibration image with an image dimension size of 64*64*3.
[0054] Pre-training: First, remove the labels of the CWRU training set data and the test set data as the data of the pre-training module, and use the data enhancement to obtain positive and negative sample pairs. Self-supervised training is performed through the BYOL comparative learning network architecture. After 150 epochs, the learning rate is 0.00003, and the batch size is 128. The trained network has considerable feature extraction capabilities. Finally, ten labeled samples are used to pre-train a closed set classifier to give it preliminary classification capabilities.
[0055] Model fine-tuning: At this stage, some unknown class data is needed. For the CWRU dataset, the normal class of the American Machinery Fault Prevention Technology Society dataset (MFTP) is used as the unknown class. The same 10 labeled samples as those used in pre-training are input into the OSCL semi-supervised learning network for training. The epoch is 128, the learning rate is 0.001, and the batch size is 64. After training, a model with good bearing fault diagnosis performance is obtained.
[0056] Test: The processed CWRU test set data is directly used as the input of the closed set classifier, and the test result is 0.9654, which is higher than the accuracy of SSCL, BYOL, Flexmatch and HSFD with the same label rate, proving the effectiveness of the method.
[0057] Specifically, the method of preprocessing the original CWRU bearing vibration signal to obtain the RGB color vibration image includes:
[0058] Before inputting into the model, the original vibration signal is preprocessed. A single vibration signal time-frequency analysis technology is difficult to cover the characteristics of a fault. At the same time, the time-frequency analysis technology describes the trend of the signal spectrum content over time, which may cause the information of the original time domain signal to be lost and incomplete. Therefore, in order to enrich the data features and improve the feature learning ability of the model, the vibration signal is converted into an RGB vibration image. Figure 2 As shown in the figure, the original vibration signal data is segmented into overlapping data of length 1024 to obtain n vibration signal samples with a length of 1024. Then, Time Domain Conversion, short-time Fourier transform (STFT) and continuous wavelet transform (CWT) are performed on each sample to obtain three two-dimensional grayscale images. The obtained two-dimensional grayscale images are merged to generate an RGB color vibration image as the input data of the model.
[0059] Specifically, the method for establishing a bearing fault diagnosis model based on RGB color vibration images includes: in the initial stage of training, the present invention uses a BYOL self-supervised contrastive learning network to pre-train an effective bearing data feature extractor. Specifically, Figure 3 As shown in Figure 1, the RGB color vibration image labels of the known class training set and test set are input into the BYOL network, and the enhanced view is obtained by randomly selecting two data augmentation methods (rotation, scaling or adding Gaussian noise) t(x) and t′(x) and
[0060] t′(x) (the vibration image after data enhancement). Input v into the online network (such as Figure 3 As shown in Figure 2, the image is processed by the ResNet50 network as the feature extractor f and the MLP network as the projector g to obtain the representation of the vibration image after data enhancement. and projection Similarly, v′ is input into the target network (such as Figure 3 As shown), it is characterized and Input the output z of the online network into the MLP network as the predictor q to get the predicted value q θ (z)(q(·) represents the Preduction prediction layer, q θ (z) refers to the output of the prediction layer), and θ represents the parameters of the online network. Figure 3 As shown, y and y' are the outputs of the intermediate layers, z' is the final output of the target network, and sg(z') represents stop-gradient, that is, the gradient of the input is not calculated (the previous layer parameters are not updated), because the target network parameter update depends on the online network (such as formula 2), because it is self-supervised training, sg(z') and q are obtained θ (z) Calculate the mean square error to make the features between the positive sample pairs more and more similar, so that the network can learn the information in the data and obtain the ability to extract features. Then, calculate q θ Similarity loss between (z) and sg(z′), sg represents the stop gradient calculation, that is, the target network parameters ξ are not updated by calculating the gradient descent. The mean square error between the online network and the target network output is The calculation is shown in Formula 1:
[0061]
[0062] During the training process, the online network iteratively updates the parameter θ through gradient descent. The parameter ξ of the target network is the exponential moving average of the parameter θ. The specific process is as shown in Formula 2. refers to the total loss of training, and The computation is the same, by feeding the enhanced view v into the target network and v′ into the online network. After training, everything except the feature extractor is discarded and used as the image feature representation.
[0063]
[0064] ξ←τξ+(1-τ)θ
[0065] Where η is the learning rate, It is calculated by inputting v into the target network and v′ into the online network, τ∈[0,1] is the target decay rate, Represents minimization via gradient descent Train the network and optimize the parameters.
[0066] After the training is completed, the feature extractor f′ of the target network is retained by contrastive learning for subsequent training to extract data features of vibration images.
[0067] In the second stage of training, the present invention uses a small amount of labeled known class training set data to supervise the pre-training of a closed set classifier. The specific training process is as follows Figure 4 As shown in the figure, the parameters of the pre-trained feature extractor are frozen during the training process (the purpose of freezing is to obtain the pre-trained feature extractor in the first stage. Self-supervised training using unlabeled data enables the contrastive learning (BYOL) network to learn the information in the fault diagnosis data and have the ability to extract features. In the second stage, it is mainly to pre-train a closed set classifier. If the parameters of the feature extractor trained in the first stage are updated at the same time, the training speed will be slowed down. In the final fine-tuning stage, the parameters of the feature extractor and the closed set classifier will continue to be optimized). The structure of the closed set classifier is two layers of fully connected layers, and the loss function is the cross entropy loss. The closed set classifier is used to classify known bearing fault types. Taking the CWRU dataset as an example, there are three main fault types of single-point defect damage: rolling element fault (BF), inner race fault (IF) and outer race fault (OF). Each fault contains three fault widths of 0.07in., 0.14in., and 0.21in. Therefore, the dataset has a total of 10 healthy states, 1 normal state (N), and 9 fault states, as shown in Table 1. A small amount of labeled data is selected from the processed CWRU dataset and input into the first stage pre-trained feature extractor in turn. The output features are input into the closed set classifier to obtain the probability of each fault type, and the highest probability is taken as the prediction result of the closed set classifier. The prediction result is compared with the corresponding true label, and the cross entropy loss between the two is calculated. By minimizing the loss and optimizing the model target, the pre-trained closed set classifier is obtained.
[0068] Table 1 CWRU data division
[0069]
[0070] Specifically, the method for fine-tuning the bearing fault diagnosis model includes:
[0071] In the final stage of training, the present invention uses a pre-trained feature extractor and closed set classifier, freezes the parameters of the feature extractor, and trains the closed set classifier and the open set classifier with a small amount of labeled known class data and a large amount of unlabeled known class and unknown class data to improve the performance of the classifier.
[0072] The closed set classifier is used as a closed set fault diagnosis classifier for closed set fault intelligent diagnosis, and the open set classifier is used to process the open set bearing fault diagnosis task. Taking the CWRU data set as an example, fault categories other than the known 10 categories are mixed into the CWRU data set. Regardless of the type of the mixed category, it is treated as the 11th category to simulate the open set bearing fault diagnosis task. The data under the open set setting is input into the open set classifier, and the classifier will output the probabilities of eleven categories. The maximum is taken as the final classification result. If it is the first 10 categories of the data, the specific bearing fault is determined by referring to the CWRU data fault type. Otherwise, it means that an unknown fault type has occurred.
[0073] Specific structure such as Figure 1 As shown in the figure, the labeled vibration image is weakly enhanced to obtain the weakly enhanced sample. The weakly enhanced sample is the vibration image after horizontal flipping or adding Gaussian noise. After the feature extractor, the closed set classifier and the multiple binary classifier are supervised and trained respectively. The closed set classifier has a supervised loss. and multiple binary classifier losses A multiple binary classifier refers to a collection of multiple binary classifiers. If the number of known classes in the data set is n, there are n binary classifiers. The result of the kth binary classifier represents whether the data belongs to the kth class. If it does not belong to any class, it is determined to be an unknown class. Strong and weak enhancements are performed on the unlabeled data to obtain strong and weak enhancement samples. The strong enhancement samples are vibration images that have been randomly cropped or have random image attributes. The weak enhancement samples are input into the closed set classifier and the multiple binary classifiers respectively. The outputs of the two classifiers are combined to obtain the pseudo labels of the unlabeled open set data. The specific formula 3 is shown as follows:
[0074]
[0075] in, Represents the probability that the unlabeled data belongs to the kth class, and the class with the largest probability is selected as the pseudo label of the open set data.
[0076] is the closed set classifier output, is the output of the multiple binary classifier, k represents the kth class, and when k=K+1, it represents an unknown class.
[0077] The strongly enhanced samples are input into the open set classifier after passing through the feature extractor, and the obtained pseudo labels are used as the target of supervised learning to obtain the loss of the open set classifier. At the same time, the strongly enhanced samples are input into the closed set classifier, and a double filtering strategy is used to select high-quality pseudo labels of known classes to optimize the performance of the closed set classifier and obtain the unlabeled known class loss. The total loss of the model is L overall , as shown in Formula 4:
[0078]
[0079] Among them, y i represents the data label, B represents the number of labeled data, μB represents the number of unlabeled data, and p i represents the output of the closed set classifier, and represent labeled and unlabeled samples respectively, o i is the output of the multiple binary classifier, is the k-th multiple binary classifier output, Respectively represent the samples processed by the weak enhancement and strong enhancement methods and input to the pseudo label, u i represents the i-th unlabeled data, τ q is the confidence threshold, represents the loss of supervised learning of the closed set classifier, The loss of the multiple binary classifiers, represents the loss of the open set classifier, Substitute the unlabeled known class data for the loss of the closed set classifier, is the total loss of the method, H(,·,) represents the cross entropy loss, 1(·) is the indicator function, F(·) is the filter function, λ mb ,λ op ,λ ui Represents the weight of the corresponding loss.
[0080] In order to verify the performance of the method proposed in the present invention under different data sets and different amounts of labeled sample data, the present invention conducts experiments on three data sets: CWRU, MFPT and XJTU-SY. At the same time, in order to verify the effectiveness of the method, the present invention is compared with Flexmatch, SSCL, BYOL and HSFD methods. The experimental results are shown in Table 2 below.
[0081] Table 2 Performance comparison of models under different datasets
[0082]
[0083]
[0084] From the table, we can see that OSCL outperforms the other four methods in most cases, especially when the labeled samples are scarce. As the number of labeled samples increases, the performance gradually improves. When there are sufficient labeled samples, all algorithms can achieve good diagnostic results.
[0085] As can be seen from the table, on the CWRU dataset, when the number of labeled samples is 10, the fault diagnosis accuracy of OSCL can reach 96.54%, but the highest accuracy of the other four methods is only 66.73%. When the number of labeled samples is 2000, the accuracy of OSCL is slightly lower than that of SSCL, but it is better than the other four methods in terms of overall performance, especially when the number of labeled samples is scarce.
[0086] The performance of the model on the MFPT dataset is similar to that on the CWRU dataset. When the amount of labeled sample data is small, such as when the number of labeled samples is 9, the accuracy of the model is more than 4.6 percentage points higher than the highest accuracy among Flexmatch, SSCL, BYOL, and HSFD. When the number of labels increases to 600, the accuracy of SSCL is higher than that of OSCL, but only 0.25 percentage points higher, which is a very small difference. Therefore, on the MFPT dataset, OSCL's overall performance is better than the other four methods.
[0087] Since the authors of the SSCL and HSFD methods did not provide source code, and there are no test results on XJTU in the paper, this paper only compares BYOL and Flexmatch on the XJTU dataset. As can be seen from the figure, except when the number of labeled samples is 300, the accuracy of Flexmatch is higher than OSCL, in other cases the OSCL method is better than the other two methods, especially when the number of labeled samples is small.
[0088] At the same time, in order to test the generalization ability of OSCL, the present invention also conducts experiments under OOD settings. As can be seen from the table, when the number of labels is scarce on the CWRU and XJTU datasets, the performance of OSCL-OOD is still better than the other three methods. It performs poorly on the MFPT dataset, but still has a high success rate. The experimental results are shown in Table 3. The experiment proves that OSCL has good generalization ability under OOD settings.
[0089] Table 3 Experimental results under open set setting
[0090]
[0091] In order to verify the effectiveness of BYOL contrastive learning pre-training feature extractor stage and IOMatch fine-tuning stage in OSCL, the present invention conducts ablation experiments on the method. The experimental results are shown in Table 4. The present invention uses CWRU data with 100 labeled samples, MFPT data with 30 samples, and XJTU data with 150 samples for experiments.
[0092] From the experimental results, it can be seen that the fine-tuning stage has a greater impact on the MFPT dataset and a smaller impact on the other two datasets. The pre-trained feature extractor stage plays an important role in the results of each dataset. Without this stage, the accuracy of the three datasets has dropped by more than 15 percentage points. In summary, both the pre-trained feature extractor stage and the model fine-tuning stage play a role in improving the model fault diagnosis performance, among which the pre-trained feature extractor stage plays a greater role.
[0093] Table 4 Ablation experiment
[0094]
[0095] In order to verify the model's feature learning ability and classification task performance for bearing fault data, the present invention uses the t-SNE algorithm to perform feature visualization on the original data and model output data respectively. Figure 5 The feature representations of the CWRU, MFPT, and XJTU datasets before and after the model is input are shown respectively. The right side is the original data representation, and the left side is the model output feature representation.
[0096] from Figure 5 It can be seen intuitively that the original data is disorganized and there is no clear boundary between each category. However, the data processed by the model can clearly distinguish each category. This shows that OSCL can well learn the characteristics of fault diagnosis data and effectively complete the bearing fault diagnosis task.
[0097] Embodiment 2
[0098] The present invention also provides a bearing fault diagnosis system based on semi-supervised contrast learning, comprising: an acquisition module, a preprocessing module, a training module, a fine-tuning module and a diagnosis module;
[0099] The acquisition module is used to acquire the original CWRU bearing vibration signal, and divide the original CWRU bearing vibration signal into a training set and a test set, wherein the training set includes: a known class and an unknown class;
[0100] The preprocessing module is used to preprocess the original CWRU bearing vibration signal to obtain an RGB color vibration image;
[0101] The training module is used to establish a bearing fault diagnosis model based on the RGB color vibration image;
[0102] The fine-tuning module is used to fine-tune the bearing fault diagnosis model;
[0103] The diagnosis module is used to perform fault diagnosis on the bearing to be tested based on the fine-tuned bearing fault diagnosis model.
[0104] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A bearing fault diagnosis method based on semi-supervised contrastive learning, characterized in that: The following steps are involved: Collecting original CWRU bearing vibration signals, and dividing the original CWRU bearing vibration signals into a training set and a test set, wherein the training set includes: a known class and an unknown class; Preprocessing the original CWRU bearing vibration signal to obtain an RGB color vibration image; Based on the RGB color vibration image, a bearing fault diagnosis model is established; Fine-tuning the bearing fault diagnosis model; Fault diagnosis is performed on the bearing to be tested based on the fine-tuned bearing fault diagnosis model.
2. The bearing fault diagnosis method based on semi-supervised contrastive learning according to claim 1 is characterized in that: The method of preprocessing the original CWRU bearing vibration signal to obtain an RGB color vibration image includes: The original CWRU bearing vibration signal data is divided into data with a length of 1024 and covered, to obtain n vibration signal samples with a length of 1024; Time Domain Conversion, STFT and CWT are performed on each vibration signal sample to obtain three two-dimensional grayscale images; The three obtained 2D grayscale images are merged to generate an RGB color vibration image.
3. The bearing fault diagnosis method based on semi-supervised contrastive learning according to claim 1 is characterized in that: Based on the RGB color vibration image, the method for establishing a bearing fault diagnosis model includes: two stages; Phase 1: Use BYOL self-supervised contrastive learning network to pre-train an effective bearing data feature extractor; The second stage: Use a small amount of labeled training set data of known classes to supervise the pre-training of a closed-set classifier.
4. The bearing fault diagnosis method based on semi-supervised contrastive learning according to claim 3 is characterized in that: The method of pre-training an effective bearing data feature extractor using BYOL self-supervised contrastive learning network includes: The RGB color vibration image labels of the known class training set and test set are input into the BYOL network, and the enhanced views v and v′ are obtained after two data augmentations; Input v into the online network, and pass through the ResNet50 network as a feature extractor and the MLP network as a projector in turn to obtain the representation y and projection z corresponding to the vibration image after data enhancement processing; Similarly, v ′ Input to the target network to get the representation y ′ and z ′ ; Input the projection z into the MLP network as a predictor to get the predicted value q θ (z).
5. The bearing fault diagnosis method based on semi-supervised contrastive learning according to claim 3 is characterized in that: Methods for supervised pre-training of a closed-set classifier using a small amount of labeled training data of known classes include: Freeze the parameters of the pre-trained bearing data feature extractor during training; The processed CWRU data set is input into the closed set classifier to obtain the probability of each extracted fault type, and the highest probability is taken as the classification result of the closed set classifier.
6. The bearing fault diagnosis method based on semi-supervised contrastive learning according to claim 1, characterized in that: The method for fine-tuning the bearing fault diagnosis model includes: In the final stage of training, the pre-trained feature extractor and closed set classifier are used, the parameters of the feature extractor are frozen, and the closed set classifier and open set classifier are trained with a preset amount of labeled known class data and a preset amount of unlabeled known class and unknown class data.
7. A bearing fault diagnosis system based on semi-supervised contrastive learning, characterized in that: include: Acquisition module, preprocessing module, training module, fine-tuning module and diagnosis module; The acquisition module is used to acquire the original CWRU bearing vibration signal and divide the original CWRU bearing vibration signal into a training set and a test set, wherein the training set includes: a known class and an unknown class; The preprocessing module is used to preprocess the original CWRU bearing vibration signal to obtain an RGB color vibration image; The training module is used to establish a bearing fault diagnosis model based on the RGB color vibration image; The fine-tuning module is used to fine-tune the bearing fault diagnosis model; The diagnosis module is used to perform fault diagnosis on the bearing to be tested based on the fine-tuned bearing fault diagnosis model.
Citation Information
Patent Citations
Bearing fault diagnosis method based on semi-supervised generative adversarial network
CN110617966A
Pump mechanical equipment bearing fault diagnosis method based on unsupervised transfer learning
CN114034486A
Electromagnetic signal individual identification method by using open set semi-supervised learning technology
CN116796167A
New fault and composite fault collaborative diagnosis method and device and storage medium
CN117171666A
Rolling bearing multi-task small sample fault diagnosis method based on semi-supervised learning
CN117743929A
Cited By
Boundary-guided rolling bearing semi-supervised fault diagnosis method and system
CN121051573A
