A rolling bearing fault diagnosis method based on semi-supervised learning

Abstract

This invention provides a semi-supervised learning-based method for rolling bearing fault diagnosis. This method introduces the concept of semi-supervised learning, utilizing contrastive learning to obtain the surface features of all unlabeled samples. Subsequently, it uses labeled samples for fine-tuning and employs a multi-label composite fault classifier to mine the latent relationships between labels and output prediction results. This method decouples labeled and unlabeled samples, allowing for flexible use of existing labels to identify new composite fault types according to the requirements of the classification task, without adding new fault categories and retraining, thus improving the model's scalability and generality.

Description

Technical Field

[0001] This invention relates to the field of subway train fault diagnosis, and in particular to a rolling bearing fault diagnosis method based on semi-supervised learning. Background Technology

[0002] During the operation of rolling bearings, in order to detect faults in a timely manner, a series of operations are required, including signal preprocessing, feature extraction, and diagnostic model construction of the collected vibration signals. With the rapid development of artificial intelligence, deep learning technology, which combines efficiency and accuracy, is widely used in the field of fault diagnosis.

[0003] Deep learning is based on "big data," enabling it to "see widely and understand deeply." Having sufficient state monitoring data is a prerequisite for fault diagnosis of mechanical systems using Convolutional Neural Networks (CNNs). Modern testing techniques make acquiring large amounts of state monitoring data easy; however, labeling this data relies on expert knowledge, is costly, and it's difficult to obtain a sufficient quantity and quality of labeled data. Typically, what's available is a large amount of unlabeled sample data. Therefore, more and more researchers are exploring how to utilize large amounts of unlabeled samples to improve classifier performance. Semi-supervised learning (SSL) is a learning model between supervised and unsupervised learning, capable of simultaneously utilizing a small number of labeled samples and a large number of unlabeled samples to achieve better classification performance. Therefore, fully utilizing unlabeled data samples and researching CNN fault diagnosis methods based on semi-supervised learning can allow intelligent fault diagnosis algorithms to be deployed in more scenarios.

[0004] Semi-supervised learning can improve the problem of poor generalization performance and decreased prediction accuracy caused by insufficient labeled samples. In current semi-supervised learning-based fault diagnosis, the distributions of most labeled and unlabeled samples are very similar. However, due to the harsh operating conditions of rolling bearings, the collected unlabeled sample data may contain noise interference. Therefore, when using semi-supervised learning for fault diagnosis, it is also necessary to pay attention to the model's performance when both labeled and unlabeled samples contain a certain range of noise intensity. Furthermore, the increasing complexity and size of mechanical systems leads to the coexistence of single and compound faults. Based on current research, there is limited research on compound fault diagnosis using semi-supervised learning. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention, based on a lightweight network that is balanced in all aspects, and considering the relatively few fault classification tasks in rolling bearing classification, adjusts the classifier structure and removes redundant dimensionality increase / decrease operations, thereby reducing model size and parameter count while ensuring classification effectiveness. To address the shortage of labeled samples, a semi-supervised learning approach is introduced to fully utilize unlabeled samples to improve the performance of existing models. To address the issue that existing models need to create new classification categories and retrain when identifying new composite fault types, a multi-label classifier is used to mine the potential relationships between labels, further enhancing multi-label classification capabilities and using existing labels to represent composite fault types.

[0006] The present invention specifically adopts the following technical solution:

[0007] A semi-supervised learning-based method for diagnosing rolling bearing faults includes the following steps:

[0008] (1) Construct a composite fault dataset for rolling bearings. Two different operating conditions with different speeds are used. The fault types are set as: normal speed, inner ring fault, outer ring fault, rolling element fault and composite fault. The composite fault includes outer ring fault and inner ring fault. Multi-label classification is performed on all fault types under the above different operating conditions to form a dataset.

[0009] (2) Unlabeled sample feature extraction: Contrastive learning is used to extract surface features from all unlabeled samples. In each round of training, random Gaussian noise is added to the one-dimensional vibration time domain signal of the rolling bearing to generate positive and negative samples. By learning the similarity among them, the maximum common surface features are obtained.

[0010] (3) A multi-label classifier is used to predict the output results. For a small number of labeled samples, a labeled sample is input to obtain the surface feature vector. After the feature vector is recombined, it is used together with the maximum common surface feature as the input signal of the multi-label classifier Transformer. All feature vectors in the input signal are encoded in terms of position and label. A fully connected layer network is used for each classification label as the output of the prediction result.

[0011] (4) Verify the performance of composite faults. Compare the output results of step (3) with the real samples, and calculate the precision, recall, mean precision and mean mean precision of the output results to verify the performance of composite faults.

[0012] Preferably, step (2) specifically includes the following steps:

[0013] Random Gaussian noise is added to the unlabeled samples of the one-dimensional vibration time-domain signal of the rolling bearing to generate a pair of positive and negative samples x. i and x j Then, the sample xi and x j Each component passes through an encoder to obtain its surface feature h. i and h j Then, it passes through a multilayer perceptron network to obtain the corresponding projection vector z. i and z j The projection vector z is maximized using a normalized temperature-scale cross-entropy loss function. i and z j The similarity; under certain rounds, as the loss function value decreases, the model gradually learns the surface features of unlabeled samples. The specific beneficial effects of this invention are as follows:

[0014] This invention utilizes contrastive learning to obtain the surface features of all unlabeled samples, followed by fine-tuning using labeled samples, and employs a classifier for multi-labeled composite faults. This method decouples labeled and unlabeled samples, allowing for flexible use of existing labels to identify new composite fault types according to the requirements of the classification task, without adding new fault categories and retraining, thus improving the model's scalability and generality. In the dataset, with 2dB of noise added, diagnostic accuracies of 76.18% and 91.65% were achieved when the labeled to unlabeled sample ratio (LTR) was 1% and 8%, respectively. In the noise robustness test, with an LTR of 8%, 2dB noise was used for the training set, while 4dB and -2dB noise were used for the test set, achieving accuracies of 91.81% and 87.35%, respectively, demonstrating insensitivity to noise. Attached Figure Description

[0015] Figure 1 This is a basic structural diagram of a rolling bearing.

[0016] Figure 2 These are diagrams illustrating several failure modes of rolling bearings.

[0017] Figure 3 This is a diagram showing the differences between supervised learning and contrastive learning.

[0018] Figure 4 This is a flowchart of the feature extraction process for unlabeled samples.

[0019] Figure 5 This is a flowchart of the semi-supervised composite fault diagnosis algorithm of the present invention.

[0020] Figure 6 Radar chart comparing the performance of three classifiers.

[0021] Figure 7 Loss function curve for noise resistance test. Detailed Implementation

[0022] As one of the most frequently used components in rotating machinery, rolling bearings are used for power transmission and supporting rotating mechanical bodies. Due to their complex and harsh working environment and continuous high-intensity operation, rolling bearings are highly prone to failure. Therefore, before conducting in-depth research on bearing fault diagnosis based on convolutional neural networks, it is necessary to understand the fault characteristics and failure causes of rolling bearings, the basic components of convolutional neural networks, and the fault datasets used in subsequent design experiments.

[0023] A rolling bearing is a precision mechanical component that transforms the sliding friction between a rotating shaft and its housing into rolling friction, thereby reducing frictional losses. A rolling bearing generally consists of four parts: an outer ring, an inner ring, rolling elements, and a cage. The inner and outer rings typically provide support, while the shape, size, and number of the rolling elements directly determine the load-bearing capacity. The cage prevents the rolling elements from falling out. (See structural diagram below.) Figure 1 As shown.

[0024] There are many reasons for rolling bearing failure, mainly including: improper assembly, poor lubrication, overload, impact, vibration, intrusion of abrasives or harmful liquids, excessively high or low ambient temperature, material defects, and low manufacturing precision. Due to the complex working environment and continuous high-intensity operation of rolling bearings, two or more failure mechanisms may act simultaneously, manifesting as a single failure mode in some cases and a combination of multiple failure modes in others. According to ISO 15234, the basic failure modes of rolling bearings during operation include:

[0025] (1) Fatigue spalling: When rolling bearings are subjected to alternating loads, the most likely place for fine cracks to appear is where the shear stress between the rolling elements and the raceway is the greatest. Over time, these fine cracks may develop into flaking, forming many small pit-like spallings, eventually causing the bearing to fail to operate normally. This phenomenon is called fatigue spalling, and it is one of the most common failure modes in rolling bearings.

[0026] (2) Wear: In rolling bearings, when the components roll or slide relative to each other, the material on their surfaces is inevitably carried away; this phenomenon is called wear. Wear leads to an increase in the clearance between components, affecting the fitting accuracy. If the lubrication conditions are poor or external debris is mixed in, the wear on the bearing surface will be more severe, generating noise and impact vibration. In severe cases, the bearing may lose its normal working ability, resulting in wear failure.

[0027] (3) Corrosion: Based on the different corrosion principles, corrosion failure can be divided into chemical corrosion and other types of corrosion. Chemical corrosion is caused by a chemical reaction triggered by moisture or other corrosive media entering the installation gap of bearing components.

[0028] (4) Plastic deformation: When a bearing is subjected to large loads or high temperatures, the local stress on the bearing surface exceeds the yield limit of the material, resulting in pits or overall deformation on the surface. If the plastic deformation reaches a certain level, it will lead to the complete failure of the bearing.

[0029] (5) Fracture and cracking: Rolling bearings may develop cracks if subjected to overload or have internal defects during normal operation. These cracks propagate as the load increases. When the load reaches the bearing's strength limit, even small cracks can cause the bearing to fracture.

[0030] When a localized fault occurs during the operation of a rolling bearing, it can cause instantaneous impacts between the contact points. Each impact excites the bearing and the entire system to undergo high-frequency free-dampening vibrations at their natural frequencies. The frequency of these instantaneous impacts (the fault characteristic frequency) depends on the bearing's rotational speed, and the impact amplitude depends on the size of the bearing fault. Figure 2 These are several failure modes of rolling bearings.

[0031] The characteristic frequency of bearing failure can be obtained by the following formula:

[0032] bearing rotation frequency f r :

[0033]

[0034] The frequency at which the rolling element passes the outer race fault point, i.e., the outer race fault characteristic frequency f. o :

[0035]

[0036] The frequency at which the rolling element passes through the inner race fault point, i.e., the inner race fault characteristic frequency f. i :

[0037]

[0038] The frequency at which damage occurs at a point on the rolling element through the inner and outer races, i.e., the characteristic frequency f of rolling element failure. e :

[0039]

[0040] The failure characteristic frequency f of the cage c :

[0041]

[0042] In the above formula, the fixed rotational speed of the bearing is n (r / min), the diameter of the rolling elements is d (mm), the pitch diameter of the bearing is D (mm), the number of rolling elements is Z, and the contact angle is α. However, in reality, factors such as bearing wear, assembly precision, and manufacturing processes can cause the actual failure characteristic frequency to differ from the calculated theoretical value. In most cases, it is necessary for analysts to manually find the failure frequency of rolling bearings close to the theoretical value.

[0043] In actual industrial processes, mechanical equipment is often a complex system, and faults tend to occur randomly and concurrently. Diagnosing single faults presents significant limitations. Furthermore, the vibration signals used for fault diagnosis have complex frequency components, making manual annotation costly and hindering the creation of a large number of labeled composite fault samples. Under these circumstances, diagnosing composite faults becomes very difficult.

[0044] To address the above issues, this invention utilizes a semi-supervised learning and multi-label classification method, employing a composite faulty rolling bearing dataset for performance validation and model robustness testing. Multi-label classification performance is validated under varying ratios of labeled to unlabeled samples (LTR).

[0045] (3) Constructing a composite fault dataset for rolling bearings

[0046] The data used in this invention is a composite fault dataset. Two operating speeds were employed: 9.9 km / h and 39.9 km / h. The fault types were: normal, inner fault, outer fault, roller fault, and composite fault. Composite faults consist of both outer and inner faults. The dataset partitioning is shown in Table 1.

[0047] Table 1 shows the percentage of labeled samples.

[0048]

[0049] Because multi-label classification is used, the fault type encoding length of this invention only requires 5 bits to represent all 10 fault types. A total of 2000 fault samples were sampled, with 100 samples from each type. The dataset was divided into four operating conditions according to the proportion of labeled samples to the total samples (LTR): 0.1%, 0.5%, 1%, 2%, and 4%.

[0050] This invention uses two multi-label classification methods for performance evaluation: Binary Relevance (BR) and Label Powerset. The Binary Relevance algorithm transforms the multi-label learning task into a fixed number of independent binary classification tasks, and its encoding form is shown in Table 2.

[0051] Table 2 BR Category Labels

[0052]

[0053]

[0054] Table 3 Label Powerset Category Tags

[0055]

[0056] Note: Labels marked with an asterisk (*) indicate sample types that do not exist in the actual dataset.

[0057] Table 3 shows the classification method using Label Powerset. Label Powerset transforms a multi-label classification problem into a single-label classification problem; that is, for a learning problem with n multi-labels, it can be transformed into at most 2n ... n This is a multi-class classification problem involving combinations of labels (the specific problem varies depending on the actual dataset classification, and different labels may have certain relationships, so not every case will occur). Each label has only two possible results: 0 and 1. In this experiment, there are 5 original labels: normal rotation speed, inner race fault, outer race fault, rolling element fault, and combined fault (inner race + outer race). After encoding with Label Powerset, excluding impossible labels, a total of 24 independent categories are formed.

[0058] (2) Feature extraction of unlabeled samples

[0059] The key to semi-supervised learning is leveraging unlabeled samples to improve model performance. Contrastive learning can provide a wealth of useful, task-irrelevant surface features from the samples. A network trained using contrastive learning differs from a typical classification network in that it does not require the model to reconstruct the original input, but rather expects the model to distinguish between different input signals (data augmentation) in the feature space. Figure 3 It concerns the differences between contrastive learning and supervised learning in the training process. From... Figure 3As shown in a and b, both supervised learning and contrastive learning utilize convolutional networks for feature extraction. However, the main differences between the two are: 1. Supervised learning uses fully connected layers (fc) after the feature extraction layer, while contrastive learning uses projection layers (proj), which are composed of multilayer perceptrons; 2. In supervised learning, the loss function mainly uses the cross-entropy loss function to measure the distance between the predicted result and the label, while in contrastive learning, the contrastive loss function is used to maximize the similarity between samples.

[0060] For this diagnostic model, the data augmentation method is to add noise to the one-dimensional vibration time-domain signal of the rolling bearing using a certain range of random Gaussian noise.

[0061] This invention uses contrastive learning to extract surface features from all unlabeled samples. Specifically, in each training round, random data augmentation is performed on the one-dimensional vibration time-domain signal of each rolling bearing to generate positive and negative samples. By learning the similarity between these samples, the maximum common surface feature is obtained. Gaussian noise of varying intensities is randomly used as the random data augmentation method for the samples. Figure 4 This is the structure of the method. x1 and x2 represent the input samples after random data augmentation of the original signal. f(.) represents the extraction of the representation vector from the augmented data sample, i.e., the model excluding the last linear layer in the convolutional network. This represents the output after the average pooling layer. g(·) is the projection layer, whose purpose is to project h. i Projected into the contrastive learning space so that the contrastive learning loss function can be used. g(·) is a multilayer perceptron network (MLP) with only 2 hidden layers. From the definition of MLP, z i =g(h i ) = W (2) σ(W (1) h i ), where σ is any activation function, such as ReLU.

[0062] Normalized temperature-scale cross-entropy NT-Xent is used as the loss function for contrastive learning, as shown in Equation 1. That is, z i and z j The regularized cosine angle. τ is the temperature coefficient. The loss function is minimized through contrastive learning-based pre-training on an unlabeled sample set. The model can effectively utilize unlabeled sample data to learn its surface features.

[0063]

[0064] (3) Use a multi-label classifier to predict the output results

[0065] The Transformer is a neural network based on a self-attention mechanism. It treats data as a sequence: data first undergoes embedding to obtain a primary representation; then, positional encoding is applied to data at corresponding positions in the sequence; finally, a learnable linear mapping maps the data to three sets of vectors: query, key, and value. The relationship between these three sets of vectors is represented through a self-attention mechanism. The formula for self-attention is:

[0066]

[0067] In image processing, compared to convolutional neural networks (CNNs), visual Transformers lack inductive bias in image data. CNNs, with their parameter-sharing mechanism, can extract local information from images and gradually increase their receptive field to integrate this local information. In contrast, visual Transformers focus more on global information. Therefore, with limited training data, visual Transformers perform worse than CNNs. This invention proposes using Transformers as multi-label classifiers to avoid the performance disadvantage of Transformers lagging behind CNNs due to insufficient training samples. In semi-supervised training, the number of unlabeled samples is often much larger than that of labeled samples. First, a CNN can be used to learn from unlabeled samples, allowing the model to acquire highly discriminative inter-class surface features for the classification task. Then, the model's feature output vector is used as the input source for the Transformer. Combining the Transformer's emphasis on global information, the relationships between output features can be fully explored, ultimately improving model performance.

[0068] This invention proposes a concurrent fault diagnosis algorithm based on semi-supervised learning. The overall process of the algorithm is as follows: Figure 5 As shown. The algorithm mainly consists of two parts. The first part is for unlabeled samples. First, random data augmentation is performed on the unlabeled samples to generate a pair of positive and negative samples x. i and x j Then, this pair of samples is encoded by an encoder to obtain surface features h. i and h j Then, after passing through an MLP perceptron, the projection vector z is obtained. i and z j Maximize z using the normalized temperature-scale cross-entropy loss function (NT-Xent). i and z jThe similarity is calculated. Over a certain number of rounds, as the loss function value decreases, the model gradually learns the surface features of unlabeled samples. The second part focuses on a small number of labeled samples. Inputting a labeled sample yields a surface feature vector h. i h i After feature reorganization, the features, along with the surface features of the unlabeled samples obtained in the first part, serve as the input signal for the Transformer encoder. Based on the characteristics of the Transformer, the input feature vector is encoded in terms of both position and label, and a fully connected layer network is used as the output of the prediction result for each classification label.

[0069] (4) Experimental verification

[0070] Because multi-label classification is more complex than single-label classification, a wider variety of evaluation metrics are needed to assess the performance of multi-label models. To better describe these metrics, some basic probabilities are first introduced, as shown in Table 4.

[0071] Table 4. Description of Evaluation Indicator Symbols

[0072]

[0073] Precision represents the percentage of actual positive samples out of the results predicted as positive.

[0074]

[0075] Recall represents the percentage of correctly predicted results out of those predicted as positive.

[0076]

[0077] Average precision (AP) is calculated for the prediction results of a single class label. This invention uses the 11-point method in Pascal 2007 to calculate AP, as shown below:

[0078]

[0079] Mean average precision (MAP) represents the average AP across all class labels, as shown below. Show:

[0080]

[0081] Hamming loss is a commonly used metric for evaluating the performance of multi-label classification. Hamming loss calculates the number of misclassified labels in the data; a smaller Hamming loss value indicates a better model performance. The formula for Hamming loss is as follows:

[0082] In addition to AP and mAP mentioned above, when comprehensively evaluating multi-label classification models, it is necessary to consider CP, CR, CF1, OP, OR, and OF1 metrics. These metrics measure model performance from two perspectives: CP, CR, and CF1 are evaluation metrics based on the dataset label categories; OP, OR, and OF1 are evaluation metrics based on the overall sample. The higher the value of these metrics, the better the model's classification performance. The calculation methods for the above two types of metrics are as follows:

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] In this invention, a composite bearing dataset will be used to validate the proposed algorithm model. The dataset will be divided into labeled and unlabeled samples (LTR), and performance analysis will be performed using different encoders.

[0090] The experiment was conducted using the PyTorch deep learning framework, with an RTX 2080 Ti graphics card and an Intel 9700 CPU. The network optimizer used Adam with an initial learning rate of 0.01. The learning rate scheduler employed a warm-up cosine decay strategy, with a warm-up phase of 40 epochs, a momentum value of 0.1, and a weight decay value of 1e-4. The network was trained for a total of 100 epochs. The improved MobileNetV3 network achieved performance comparable to the original network, but with fewer parameters and the fastest prediction time. Therefore, it was used as the surface feature learner in stage 1. Different networks were used as classifiers in stage 2, with fully connected layers (FC) and convolutional layers (CNN) used for performance comparison.

[0091] The experimental results of each model under different LTR settings on the composite bearing fault dataset are shown in Tables 5, 6, and 7. As can be seen from the tables, the method proposed in this invention achieves optimal performance under all LTR conditions. The ML-SSL method proposed in this invention achieves the best results at LTRs of 0.1%, 0.5%, 1%, 2%, 4%, and 8%. As the LTR increases, the number of FC (fully connected layer) parameters decreases, limiting the range of model weight adjustments, thus its accuracy improvement is not significant. CNN and the method of this invention (using Transformer as a multi-label classifier) ​​involve adding an external network in stage 2 to further learn classification features.

[0092] Table 5 Performance verification of the method of the present invention

[0093]

[0094] Table 6 CNN Performance Validation (Stage 2 Classifier)

[0095]

[0096] Table 7 FC Performance Validation (Stage 2 Classifier)

[0097]

[0098] Since this invention uses multi-label classification for composite fault identification, the accuracy of each label is a key indicator. Tables 8, 9, and 10 show the label accuracy of each comparison network.

[0099] Table 8. Accuracy of Transformer Classifier for Each Label

[0100]

[0101] Table 9. Accuracy of CNN classifier for each label

[0102]

[0103] Table 10 Accuracy of each label for the FC classifier

[0104]

[0105]

[0106] Figure 6 This radar chart, created using data from Tables 8, 9, and 10, provides a more intuitive display of the classification. Figure 6It is evident that normal labels are the easiest to identify, while speed labels representing bearing operating conditions are the most difficult. The proposed method demonstrates superior performance in the recognition rate of all labels, particularly for speed labels representing rolling bearing operating conditions. This invention primarily utilizes multiple labels for composite faults, with the composite fault label composed of both inner ring faults and outer ring faults. The accuracy of inner and outer ring faults reflects, to some extent, the model's ability to distinguish between composite faults and single fault points.

[0107] Label Powerset is used for more rigorous multi-label model classification to demonstrate the performance of ML-SSL in identifying complex faults under strict classification conditions. Label Powerset transforms multi-label classification into a multi-classification task, resulting in some classifications not present in the actual data samples. These unpresented classifications are now merged into a single category called "No Practical Meaning" for display. Specific data is shown in Table 11.

[0108] Table 11 Experimental results of the Label Powerset classification algorithm

[0109]

[0110]

[0111] In the Label Powerset classification method, compared to binary classification, it more rigorously categorizes the classification cases. If the model strictly classifies composite faults under each operating condition, its overall accuracy is lower than that of binary classification. This is because the model outputs some label combinations that do not exist in the actual samples (meaningless label combinations). However, this method also demonstrates some advantages of using multi-labels for composite fault diagnosis. Since composite faults are caused by multiple single-point faults acting simultaneously, the diagnostic network should, even if it fails to identify the composite fault, should still identify one of the single-point faults as stably as possible.

[0112] The composite faults in this experimental dataset are caused by the combined effects of the inner and outer rings. As shown in Table 11, the biggest source of misclassification in multi-condition composite faults is the error in condition identification. For example, in the classification of composite faults at speed 1, 15 samples were incorrectly classified as composite faults at speed 2; in the classification of composite faults at speed 2, 11 samples were incorrectly classified as composite faults at speed 1.

[0113] In the statistics of misclassification of complex faults at speed 1, only 4 samples were classified as meaningless errors, while the remaining misclassifications were related to the inner and outer races, resulting in an accuracy of 97%. In the statistics of misclassification of complex faults at speed 2, 5 samples were classified as meaningless errors, 3 samples were misclassified as single rolling element faults, and the rest were related to the inner and outer races, resulting in an accuracy of 94.55%. This demonstrates that using multiple labels for complex fault identification under multiple operating conditions is more accurate in identifying single-point faults.

[0114] In the contrastive learning phase, random-intensity white noise is used to augment the unlabeled samples. Therefore, the model exhibits a certain advantage in noise robustness. To verify this conclusion, the following two experiments were conducted. The experimental procedure is as follows: 1. In the contrastive learning phase (Phase 1), random noise within a certain range is added to the unlabeled data. 2. In the phase of adjusting the model weights using labeled data (Phase 2), the model is trained using only the specified noise. 3. On the test set (Phase 3), the final performance of the model is tested using the specified noise intensity. Both experiments were conducted at LTR = 8%, and other parameters are shown in Table 12.

[0115] Table 12 Experimental parameter settings for model noise immunity test

[0116]

[0117] Tables 13 and 14 contain detailed test data for Experiment 1 and Experiment 2, respectively.

[0118] Table 13 Accuracy of Noise Resistance Test in Experiment 1

[0119]

[0120] Table 14 Accuracy of Noise Resistance Test in Experiment 2

[0121]

[0122] from Figure 7The training loss function curve for Stage 1 shows that using white noise in the range of [0dB, 4dB] for contrastive learning on unlabeled data resulted in a lower loss function value. Tables 13 and 14 show that after contrastive learning, the model can indeed find common high-discrimination features among unlabeled samples at different noise intensities. In Experiment 1, after fine-tuning the model using only 2dB of labeled sample data, good results were achieved on the test set at 4dB, 2dB, 0dB, and -2dB. While the noise intensities at 0dB and -2dB were greater than the 0dB noise used during training, the test results did not decrease significantly, decreasing by only 0.91% and 2.93% respectively compared to 2dB. The same situation occurred in Experiment 2; using 0dB noise during training, the accuracy only decreased by 0.18% and 2.38% on the -2dB and -4dB test sets. However, it is worth noting that even using 0dB and -2dB as the noise in the test set, the accuracy results of Experiments 1 and 2 differed. The main reason for this difference is the speed label of the rolling bearing operating conditions. Because identifying operating conditions is difficult, when using stronger data augmentation noise in contrastive learning, the model may be unable to find more common features with high discriminative power to distinguish speeds. However, for labels with clearly distinguishable features, such as normal and rolling labels, the model can still find their common features well even with stronger noise.

Claims

1. A rolling bearing fault diagnosis method based on semi-supervised learning, characterized in that, The method comprises the following steps: (1) Constructing a composite fault data set of rolling bearings, using two different rotating speed working condition data, setting the fault types as: normal rotating speed, inner ring fault, outer ring fault, rolling element fault and composite fault, the composite fault including outer ring fault and inner ring fault, performing multi-label classification on all fault types under the above different working conditions to form a data set; (2) Unlabeled sample feature extraction, using contrast learning to extract surface features of all unlabeled samples, in each training round, using random Gaussian noise to add noise to the one-dimensional vibration time domain signal of the rolling bearing, generating positive and negative samples, obtaining the maximum common surface feature by learning the similarity thereof; (3) Using a multi-label classifier to predict the output result, for a small number of labeled samples, inputting a labeled sample to obtain a surface feature vector, recombining the feature vector and the maximum common surface feature to jointly serve as an input signal of a multi-label classifier Transformer, performing position coding and label coding on all feature vectors in the input signal, and using a fully connected layer network for each classification label as the output of the prediction result; (4) Composite fault performance verification, comparing the output result of step (3) with the real sample to calculate the precision, recall, average precision and mean average precision of the output result, thereby verifying the composite fault performance.

2. The semi-supervised learning based rolling bearing fault diagnosis method of claim 1, wherein, Step (2) specifically comprises the following steps: A pair of positive and negative samples x i and x j are generated by adding random Gaussian noise to the unlabeled samples of one-dimensional vibration time-domain signals of rolling bearings i and x j are encoded by an encoder, respectively, to obtain surface features h i and h j , which are then input into a multi-layer perception network to obtain corresponding projection vectors z i and z j , and the similarity of the projection vectors z i and z j is maximized by using a normalized temperature scale cross-entropy loss function; under certain rounds, the model gradually learns the surface features of the unlabeled samples as the loss function value decreases.

Images