A bearing fault diagnosis method and system based on semi-supervised contrastive learning

Abstract

This invention discloses a bearing fault diagnosis method and system based on semi-supervised contrastive learning, comprising: acquiring raw CWRU bearing vibration signals and dividing the raw CWRU bearing vibration signals into a training set and a test set, wherein the training set includes known classes and unknown classes; preprocessing the raw CWRU bearing vibration signals to obtain RGB color vibration images; establishing a bearing fault diagnosis model based on the RGB color vibration images; fine-tuning the bearing fault diagnosis model; and performing fault diagnosis on the bearing under test based on the fine-tuned bearing fault diagnosis model. This invention proposes a semi-supervised learning method (OSCL) based on contrastive learning for bearing fault diagnosis, which utilizes a small amount of labeled data of known classes and a large amount of unlabeled data containing known and unknown classes for joint training, making it more suitable for intelligent bearing fault diagnosis in real industrial scenarios.

Description

Technical Field

[0001] This invention relates to vibration signal processing methods, semi-supervised learning methods, and contrastive learning methods, belonging to the field of bearing fault diagnosis, and specifically to a bearing fault diagnosis method and system based on semi-supervised contrastive learning. Background Technology

[0002] Rolling bearings are among the most widely used general-purpose mechanical components in rotating machinery, playing a crucial role. Their proper operation directly impacts the entire system's performance. The rolling relationship between the main components of a bearing makes it one of the most vulnerable parts of the machine; bearing failure can lead to high maintenance costs, significant economic losses, and safety hazards. Therefore, intelligent bearing fault diagnosis, as a vital part of fault prediction and health management, is of great significance for ensuring the safe and stable operation of rotating machinery and is a worthy research topic.

[0003] Intelligent bearing fault diagnosis primarily involves datasets of time-series signals collected by sensors, requiring extensive domain knowledge. Machine learning-based bearing fault diagnosis methods rely heavily on specialized expertise and complex feature engineering. In contrast, deep learning can directly extract and learn features from raw input data without requiring manually designed features or domain expertise. Therefore, many researchers have attempted to apply deep learning methods to intelligent bearing fault diagnosis, achieving significant success. However, deep learning-based fault diagnosis methods typically require large amounts of labeled data, but in real-world industrial scenarios, only a small portion of the acquired data is labeled. The insufficient number of labeled samples has become a significant obstacle in this field. Consequently, semi-supervised learning has gradually become a research hotspot in bearing fault diagnosis.

[0004] In recent years, researchers have proposed several semi-supervised learning methods for bearing fault diagnosis. While these methods have improved bearing fault diagnosis performance to some extent when labeled samples are scarce, they neglect the possibility of unknown fault types, i.e., outliers (OODs). In real-world industrial scenarios, fault types are diverse, and many unexpected situations can arise, leading to new fault types. These undefined fault data can become outliers in the training data, thereby reducing model performance and preventing the methods from achieving their intended results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a bearing fault diagnosis method and system based on semi-supervised contrastive learning. It introduces an open-set classifier into the traditional semi-supervised bearing fault diagnosis framework, and combines a binary classifier and a closed-set classifier to generate pseudo-labels as open-set classification targets.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A bearing fault diagnosis method based on semi-supervised contrastive learning includes the following steps:

[0008] The original CWRU bearing vibration signal is collected and divided into a training set and a test set. The training set includes known classes and unknown classes.

[0009] The original CWRU bearing vibration signal is preprocessed to obtain an RGB color vibration image;

[0010] Based on the RGB color vibration images, a bearing fault diagnosis model is established;

[0011] The bearing fault diagnosis model was fine-tuned.

[0012] Based on the finely tuned bearing fault diagnosis model, fault diagnosis is performed on the bearing under test.

[0013] Preferably, the method for preprocessing the original CWRU bearing vibration signal to obtain an RGB color vibration image includes:

[0014] The original CWRU bearing vibration signal is divided into data segments with a length of 1024, resulting in n vibration signal samples with a length of 1024.

[0015] Three two-dimensional grayscale images were obtained by performing Time Domain Conversion, STFT, and CWT on each vibration signal sample;

[0016] The three obtained two-dimensional 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 includes two stages;

[0018] Phase 1: Pre-train an effective bearing data feature extractor using the BYOL self-supervised contrastive learning network;

[0019] The second stage involves supervised pre-training of a closed-set classifier using a small amount of labeled training data of known classes.

[0020] Preferred methods for pre-training an effective bearing data feature extractor using a BYOL self-supervised contrastive learning network include:

[0021] Label-free RGB color vibration images from the known training and test sets are input into the BYOL network, and augmented views are obtained after two data augmentation processes. and ;

[0022] Will The input is fed into an online network, passing sequentially through a ResNet50 network as a feature extractor and an MLP network as a projector, to obtain the representation of the vibration image after data augmentation. and projection ;

[0023] Similarly Input is given to the target network to obtain the representation. and ;

[0024] Project The input is fed into an MLP network acting as a predictor to obtain the predicted value. .

[0025] Preferred methods for supervised pre-training of a closed-set classifier using a small amount of labeled training set data of known classes include:

[0026] Freeze the parameters of the pre-trained bearing data feature extractor during training;

[0027] The processed CWRU dataset 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, a pre-trained feature extractor and a closed-set classifier are used. The parameters of the feature extractor are frozen, and the closed-set classifier and the open-set classifier are trained with a preset amount of labeled known class data and a preset amount of unlabeled known and unknown class data.

[0030] The present invention also provides a bearing fault diagnosis system based on semi-supervised contrastive 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. The training set includes known classes and unknown classes.

[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 diagnostic module is used to diagnose the bearing under test based on the fine-tuned bearing fault diagnosis model.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] In real-world industrial scenarios, due to the low frequency of failures, the number of vibration signal failure samples collected is far smaller than the number of healthy state samples. Furthermore, the high cost of manual labeling further exacerbates this scarcity of labeled fault data. Simultaneously, a large amount of unlabeled fault data may contain unknown fault types, thus impacting diagnostic performance. To address these issues, this invention proposes a semi-supervised learning method (OSCL) based on contrastive learning for bearing fault diagnosis. This method utilizes a small amount of labeled data with known classes and a large amount of unlabeled data containing both known and unknown classes for joint training, making it more suitable for intelligent bearing fault diagnosis in real-world industrial scenarios. Attached Figure Description

[0038] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of a bearing fault diagnosis method based on semi-supervised comparative learning according to an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the data preprocessing process according to an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the feature extractor training process according to an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the training process of the closed-set classifier according to an embodiment of the present invention;

[0043] Figure 5 This is a flowchart illustrating the specific implementation steps of an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1

[0047] like Figure 1 As shown, this invention provides a bearing fault diagnosis method based on semi-supervised contrastive learning, comprising the following steps:

[0048] We collected bearing vibration signals from the original Case Western Reserve University (CWRU) dataset and divided the original CWRU bearing vibration signals into a training set and a test set. The training set included known classes and unknown classes.

[0049] The original CWRU bearing vibration signal was preprocessed to obtain an RGB color vibration image.

[0050] A bearing fault diagnosis model was established based on RGB color vibration images;

[0051] Fine-tuning of the bearing fault diagnosis model;

[0052] The bearing under test is diagnosed based on the finely tuned bearing fault diagnosis model.

[0053] The specific implementation process is as follows: Figure 5 As shown, data preparation involved dividing the original CWRU bearing vibration signal into training and test sets. The training set data was processed separately, aligning and performing CWT, STFT, and Time Domain Conversion to obtain three 2D grayscale images. These three grayscale images were then merged to obtain an RGB color vibration image with a dimension of [missing information]. .

[0054] Pre-training: First, the CWRU training and test sets were stripped of their labels and used as the pre-training data. The input data was augmented to obtain positive and negative sample pairs. Self-supervised training was then performed using a BYOL contrastive learning network architecture for 150 epochs with a learning rate of 0.00003 and a batch size of 128. The trained network exhibited considerable feature extraction capabilities. Finally, a closed-set classifier was pre-trained using ten labeled samples to provide preliminary classification ability.

[0055] Model fine-tuning: This stage requires some unknown class data. For the CWRU dataset, the normal class of the American Society for Mechanical Failure Prevention (MFTP) dataset is used as the unknown class. The same 10 labeled samples as during 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] Testing: Using the processed CWRU test set data directly as input to the closed-set classifier, the test result was 0.9654, which is higher than the accuracy of SSCL, BYOL, Flexmatch and HSFD under the same label rate, proving the effectiveness of the method.

[0057] Specifically, methods for preprocessing the original CWRU bearing vibration signal to obtain an RGB color vibration image include:

[0058] Before being input into the model, the original vibration signal undergoes data preprocessing. Time-frequency analysis of a single vibration signal is insufficient to encompass the characteristics of different faults, and since it describes the trend of signal spectral content changes over time, it may lead to the loss or incompleteness of information from the original time-domain signal. Therefore, to enrich data features and improve the model's feature learning ability, the vibration signal is converted into an RGB vibration image. For example... Figure 2 As shown, the original vibration signal data is divided into overlapping data segments of length 1024 to obtain n vibration signal samples of length 1024. Then, TimeDomain 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, which is used 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 training stage, this invention utilizes a BYOL self-supervised contrastive learning network to pre-train an effective bearing data feature extractor. Specifically, as follows... Figure 3 As shown, RGB color vibration images of known classes from the training and test sets are delabeled and input into the BYOL network. Two data augmentation methods (rotation, scaling, or adding Gaussian noise) are then randomly selected. and Get Enhanced View and (Vibration image after data augmentation). Input to an online network (such as Figure 3 As shown), it passes through the feature extractor in sequence. ResNet50 network as a projector The MLP network is used to obtain the representation of the vibration image after data augmentation. and projection Similarly, Input to the target network (e.g.) Figure 3 As shown), the characterization is obtained. and Output from the online network Input as a predictor In the MLP network, the predicted value is obtained. ( The representative is the Preduction prediction layer. (Refers to the output of the prediction layer) This represents parameters of the online network. For example... Figure 3 As shown, y and y' It is the output of the intermediate layer. z' It is the final output of the target network. This represents stop-gradient, meaning the gradient of the input is not calculated (the parameters of previous layers are not updated), because the parameter updates of the target network depend on the online network (as shown in Equation 2). Since it is self-supervised training, the obtained... and Mean squared error is calculated to make the features between positive sample pairs increasingly similar, enabling the network to learn information from the data and gain feature extraction capabilities. Then, the mean squared error is calculated. and Loss of similarity between them This signifies stopping gradient calculation, i.e., stopping the calculation of the target network parameters. Updated without calculating gradient descent. Mean squared error of the outputs of the online network and the target network. The calculation is shown in Formula 1:

[0060]

[0061] During training, the online network iteratively updates its parameters using gradient descent. Parameters of the target network It is a parameter The exponential moving average is calculated using Formula 2. This refers to the total training loss. and The calculation method is the same, by obtaining an enhanced view. Input to the target network The input is fed into an online network. After training, everything except the feature extractor is discarded, and the result is used as the image feature representation.

[0062]

[0063] in, It's the learning rate. It is by... Input the target network and The result is obtained from online network calculation. It is the target decay rate. This represents minimizing via gradient descent. Train the network and optimize the parameters.

[0064] After training, the feature extractor of the target network is learned through contrastive learning. Data features are used to extract vibration images for subsequent training.

[0065] In the second stage of training, this invention utilizes a small amount of labeled training set data of known classes to supervise the pre-training of a closed-set classifier. The specific training process is as follows: Figure 4 As shown, the parameters of the pre-trained feature extractor are frozen during training (the purpose of freezing is to obtain the pre-trained feature extractor in the first stage. Self-supervised training using unlabeled data allows the contrastive learning (BYOL) network to learn information from the fault diagnosis data and acquire feature extraction capabilities. In the second stage, it is mainly for pre-training a closed-set classifier. If the parameters of the feature extractor trained in the first stage are updated simultaneously, 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 closed-set classifier has a structure of two fully connected layers, and the loss function is cross-entropy loss. The closed-set classifier is used to classify known bearing fault types. Taking the CWRU dataset as an example, single-point defect damage has three main fault types: rolling element fault (BF), inner race fault (IF), and outer race fault (OF). Each fault includes three fault widths: 0.07 in., 0.14 in., and 0.21 in. Therefore, the dataset has 10 healthy states, 1 normal state (N), and 9 fault states, as shown in Table 1. A small number of labeled data points from the processed CWRU dataset are sequentially input into the first-stage pre-trained feature extractor. The output features are then input into a closed-set classifier to obtain the probability of each fault type. 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, the model objective is optimized to obtain the pre-trained closed-set classifier.

[0066] Table 1 CWRU Data Division

[0067]

[0068] Specifically, the method for fine-tuning the bearing fault diagnosis model includes:

[0069] In the final stage of training, this invention employs a pre-trained feature extractor and a 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 and unknown class data, thereby improving the performance of the classifier.

[0070] A closed-set classifier is used as a closed-set fault diagnosis classifier for intelligent closed-set fault diagnosis, while an open-set classifier is used to handle open-set bearing fault diagnosis tasks. Taking the CWRU dataset as an example, fault categories other than the known 10 classes are mixed into the CWRU dataset. Regardless of the type of the mixed-in categories, they are all treated as the 11th class, simulating an open-set bearing fault diagnosis task. The data under the open-set settings is input into the open-set classifier, which outputs the probabilities of eleven classes. The probability of the highest probability is taken as the final classification result. If it is among the first 10 classes in the data, the specific bearing fault is determined by referring to the fault types in the CWRU data; otherwise, it indicates an unknown fault type.

[0071] The specific structure is as follows: Figure 1 As shown, weak enhancement samples are obtained by weakly enhancing labeled vibration images. These weakly enhanced samples are vibration images that have been horizontally flipped or have Gaussian noise added. After passing through a feature extractor, a supervised training closed-set classifier and a multi-binary classifier are used to obtain the closed-set classifier with supervised loss. and the loss of the multi-binary classifier A multi-binary classifier refers to a collection of multiple binary classifiers. If the dataset has n known classes, then there are n binary classifiers. The result of the k-th binary classifier indicates whether the data belongs to the k-th class. If it doesn't belong to any class, it's classified as an unknown class. Strong and weak augmentation are performed on the unlabeled data to obtain strongly augmented and weakly augmented samples. Strongly augmented samples are vibration images with random cropping or random image attributes. The weakly augmented samples are input into the closed-set classifier and the multi-binary classifier, respectively. The outputs of the two classifiers are used together to obtain the pseudo-label for the unlabeled open-set data, as shown in Formula 3.

[0072]

[0073] in, The unlabeled data belongs to the first k The probability of each class is calculated, and the class with the highest probability is selected as the pseudo-label for the open set data.

[0074] It is the output of the closed-set classifier. It is the output of a multi-binary classifier. , k Representing the k Class, when k = K +1 indicates an unknown class.

[0075] The strongly enhanced samples are processed by a feature extractor and then fed into an open-set classifier. The resulting pseudo-labels are used as the target for supervised learning, and the loss of the open-set classifier is obtained. Simultaneously, strongly augmented samples are input into the closed-set classifier, and a dual filtering strategy is used to select high-quality pseudo-labels for known classes, optimizing the performance of the closed-set classifier and obtaining the loss for the unlabeled known classes. The total loss of the model is Specifically, see Formula 4:

[0076]

[0077] in, Represents data labels, This represents the number of labeled data items. Represents the number of unlabeled data items. The output of the closed-set classifier is represented. and These represent labeled and unlabeled samples, respectively. It is the output of a multi-binary classifier. It is the output of the k-th multi-binary classifier. These represent the input samples to the pseudo-labels after being processed by mild and strong enhancement methods, respectively. Represents the i-th unlabeled data. It is a confidence threshold. The loss represents the supervised learning loss of the closed-set classifier. The loss of the multiple binary classifier is calculated by the proxy. Represents the loss of the open set classifier. The loss of a closed-set classifier is calculated by inputting unlabeled known-class data into it. It is the total loss of the method. Represents cross-entropy loss, For index functions, It is a filter function. This represents the weight of the corresponding loss.

[0078] To verify the performance of the proposed method under different datasets and different amounts of labeled sample data, experiments were conducted on three datasets: CWRU, MFPT, and XJTU-SY. Furthermore, to verify the effectiveness of the method, it was compared with Flexmatch, SSCL, BYOL, and HSFD methods. The experimental results are shown in Table 2 below.

[0079] Table 2 Performance comparison of models on different datasets

[0080]

[0081] The table shows that OSCL outperforms the other four methods in most cases, especially when labeled samples are scarce. Performance gradually improves with increasing labeled sample count. With sufficient labeled samples, all algorithms achieve good diagnostic results.

[0082] As can be seen from the table, on the CWRU dataset, when the number of labeled samples is 10, the accuracy of OSCL fault diagnosis can reach 96.54%, while 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 its overall performance, especially when the number of labeled samples is small, is better than the other four methods.

[0083] The model's performance on the MFPT dataset is similar to its performance on the CWRU dataset. When the number of labeled samples is small, such as 9, the model's accuracy 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, SSCL's accuracy is higher than OSCL, but only by 0.25 percentage points, a negligible difference. Therefore, on the MFPT dataset, OSCL generally outperforms the other four methods.

[0084] Since the authors of SSCL and HSFD did not provide source code, and the paper does not include test results on XJTU, this invention was only compared with BYOL and Flexmatch on the XJTU dataset. As can be seen from the figure, except when the number of labeled samples is 300, Flexmatch has a higher accuracy than OSCL; in all other cases, the OSCL method outperforms the other two methods, especially when the number of labeled samples is small.

[0085] Meanwhile, to test the generalization ability of OSCL, experiments were also conducted under the OOD setting. As can be seen from the table, OSCL-OOD still outperforms the other three methods when the number of labels is sparse on the CWRU and XJTU datasets. It performs worse on the MFPT dataset, but still maintains a high success rate. The experimental results are shown in Table 3. These experiments demonstrate that OSCL has good generalization ability under the OOD setting.

[0086] Table 3. Experimental results under the open set setting

[0087]

[0088] To verify the effectiveness of the BYOL contrastive learning pre-training feature extractor stage and IOMatch fine-tuning stage in OSCL, this invention conducted ablation experiments on the method. The experimental results are shown in Table 4. This invention used CWRU data with 100 labeled samples, MFPT data with 30 samples, and XJTU data with 150 samples for the experiments.

[0089] The experimental results show that the fine-tuning stage has a greater impact on the MFPT dataset than on the other two datasets. The pre-training feature extractor stage plays a crucial role in the results for each dataset; removing this stage results in a drop in accuracy of more than 15 percentage points across all three datasets. In summary, both the pre-training feature extractor stage and the model fine-tuning stage contribute to improving the model's fault diagnosis performance, with the pre-training feature extractor stage having a greater impact.

[0090] Table 4 Ablation Experiment

[0091]

[0092] To verify the model's feature learning ability and classification performance on bearing fault data, this invention uses the t-SNE algorithm to visualize the features of the original data and the model output data. Figure 5 The images show the feature representations of the CWRU, MFPT, and XJTU datasets before and after inputting them into the model. The right side shows the original data representation, and the left side shows the model output feature representation.

[0093] from Figure 5 As can be clearly seen from this invention, the original data was disorganized, with no clear boundaries between categories. However, the data processed by the model clearly distinguishes each category. This demonstrates that OSCL can effectively learn the characteristics of fault diagnosis data and efficiently complete the bearing fault diagnosis task.

[0094] Example 2

[0095] The present invention also provides a bearing fault diagnosis system based on semi-supervised contrastive learning, comprising: an acquisition module, a preprocessing module, a training module, a fine-tuning module, and a diagnosis module;

[0096] 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. The training set includes known classes and unknown classes.

[0097] The preprocessing module is used to preprocess the original CWRU bearing vibration signal to obtain an RGB color vibration image;

[0098] The training module is used to establish a bearing fault diagnosis model based on the RGB color vibration images;

[0099] The fine-tuning module is used to fine-tune the bearing fault diagnosis model;

[0100] The diagnostic module is used to diagnose the bearing under test based on the fine-tuned bearing fault diagnosis model.

[0101] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A bearing fault diagnosis method based on semi-supervised contrastive learning, characterized in that, The method comprises the following steps: 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 comprises known classes and unknown classes; Pretreating the original CWRU bearing vibration signals to obtain RGB color vibration images; Based on the RGB color vibration images, a bearing fault diagnosis model is established; Fine-tuning the bearing fault diagnosis model; Based on the fine-tuned bearing fault diagnosis model, the bearing to be tested is diagnosed for faults; Based on the RGB color vibration images, a method for establishing a bearing fault diagnosis model comprises two stages; The first stage: an effective bearing data feature extractor is pre-trained by using a BYOL self-supervised contrast learning network; The second stage: a closed-set classifier is supervised pre-trained by using a small amount of labeled known class training set data; The method for pre-training an effective bearing data feature extractor by using a BYOL self-supervised contrast learning network comprises: The RGB color vibration image of the known class training set and test set is labeled and input into the BYOL network, and two data enhancements are obtained and ; Will be input to the online network, in turn through the ResNet50 network as a feature extractor, as a projector MLP network, get through the data enhancement processing of the vibration image corresponding to the representation and projection and projection ; Also to be input to the target network are to obtain a representation and ; projecting into an MLP network as a predictor, obtaining a predicted value ; The method for supervised pre-training a closed-set classifier by using a small amount of labeled known class training set data comprises: In the training process, the parameters of the pre-trained bearing data feature extractor are frozen; The CWRU data set processed by data is input into the closed-set classifier to obtain the probability of each type of fault, and the highest probability is taken as the classification result of the closed-set classifier; The method for fine-tuning the bearing fault diagnosis model comprises: In the last stage of training, the pre-trained feature extractor and the closed-set classifier are used, the parameters of the feature extractor are frozen, and the closed-set classifier and the open-set classifier are trained by using a preset amount of labeled known class data and a preset amount of unlabeled known class and unknown class data.

2. The bearing fault diagnosis method based on semi-supervised contrastive learning according to claim 1, characterized in that, The method for pretreating the original CWRU bearing vibration signals to obtain RGB color vibration images comprises: The original CWRU bearing vibration signals are segmented by data with a length of 1024 and overlap to obtain n vibration signal samples with a length of 1024; Each vibration signal sample is subjected to Time Domain Conversion, STFT and CWT to obtain three two-dimensional gray images; The three two-dimensional gray images obtained are combined to generate an RGB color vibration image.

3. A bearing fault diagnosis system based on semi-supervised contrastive learning, the system is used to implement the method of any one of claims 1-2, characterized in that, It comprises: A collection module, a pretreatment module, a training module, a fine-tuning module and a diagnosis module; The collection module is used to collect original CWRU bearing vibration signals, and divide the original CWRU bearing vibration signals into a training set and a test set, wherein the training set comprises known classes and unknown classes; The pretreatment module is used to pretreat the original CWRU bearing vibration signals to obtain RGB color vibration images; The training module is used to establish a bearing fault diagnosis model based on the RGB color vibration images; The fine-tuning module is used to fine-tune the bearing fault diagnosis model; The diagnosis module is used to diagnose the bearing to be tested for faults based on the fine-tuned bearing fault diagnosis model.

Images