An inter-domain data label relationship unknown rolling bearing cross-domain fault diagnosis method and system

Abstract

This invention discloses a method and system for cross-domain fault diagnosis of rolling bearings with unknown inter-domain data label relationships. It relates to the field of rolling bearing fault diagnosis technology and aims to solve the problem of low fault diagnosis accuracy in existing transfer learning methods due to unknown data label relationships between the source and target domains. The key technical points of this invention include: constructing a comparative general domain adaptation model; introducing a BYOL network to mine the unique structure of the target domain data and proposing an entropy separation strategy to reject unknown class samples; simultaneously, designing a source class weighting mechanism to improve the transfer semantic enhancement method, assigning different class-level weights to the source classes, so that the feature distributions of the two domains are better aligned in the shared label space, further constructing a fault diagnosis model; and performing fault diagnosis through the trained fault diagnosis model. This invention exhibits superior fault diagnosis accuracy under different operating conditions.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing fault diagnosis technology, and specifically to a method and system for cross-domain fault diagnosis of rolling bearings where the inter-domain data label relationship is unknown. Background Technology

[0002] Rolling bearings, as a key basic component in rotating machinery, are widely used in many fields related to the national economy and national defense. [1] In practical applications, rolling bearings operate under complex and variable conditions. A failure can range from affecting machine performance to causing major accidents and resulting in injuries or fatalities. [2] Therefore, fault diagnosis of rolling bearings under different operating conditions is of great significance. [3] .

[0003] In recent years, deep learning has been widely used in the field of intelligent fault diagnosis due to its outstanding feature extraction capabilities. Reference [4] extracts fault features through a bidirectional long short-term memory network and inputs them into a Softmax layer to complete fault classification. Reference [5] uses the firefly algorithm for parameter optimization and uses a multi-channel convolutional neural network to extract features, thus realizing the fault diagnosis of rolling bearings. Reference [6] converts the original vibration signal into a two-dimensional image and inputs it into an improved convolutional neural network, achieving good fault diagnosis results. Reference [7] uses adaptive slope blocks to improve the residual shrinkage unit and successfully applies it to fault diagnosis in a noisy background.

[0004] The aforementioned deep learning-based fault diagnosis method requires two prerequisites: a large-scale, high-quality labeled dataset, and the test data and training data coming from the same distribution. [8] However, under different operating conditions, the training and testing data for rolling bearings often exhibit significant distributional differences. Furthermore, collecting large amounts of labeled data for training is often difficult and expensive.

[0005] To address the shortcomings of deep learning-based fault diagnosis methods, transfer learning is introduced into cross-domain fault diagnosis, aiming to transfer knowledge learned from the source domain to the target domain. [9] Domain adaptation technology, as one of the most popular transfer learning methods, has received increasing attention. Reference

[10] uses multi-kernel to solve nonlinear problems and dynamically aligns edge distributions and conditional distributions, improving robustness in cross-domain fault diagnosis. Reference

[11] proposes a Gaussian-guided feature alignment strategy, achieving a high cross-domain fault diagnosis accuracy. Reference

[12] proposes an improved conditional adversarial mechanism that simultaneously adapts features and labels, successfully applying it to cross-domain bearing fault diagnosis. Reference

[13] constructs a transfer diagnosis model based on a deep attention network, realizing fault diagnosis of rolling bearings at different speeds.

[0006] The aforementioned fault diagnosis methods all employ closed-set domain adaptation, meaning that the source and target domain data share the same health status type. However, in real-world industrial scenarios, it's impossible to obtain the relationship between the source and target domain data labels in advance. The source and target domain data may have shared and private health statuses; directly aligning the feature distributions of the two domains will lead to negative transfer in the diagnostic model.

[14] .

[0007] To address the limitations of closed-set domain adaptation, general domain adaptation can correctly classify the health status shared with the source domain in the target domain when the data label relationship between domains is unknown. Reference

[15] proposes a general domain adaptation method with mixed weighting of source class and target instance to achieve class-level alignment of unknown fault modes and shared health status, and verifies the effectiveness of the proposed method on a rotating machinery dataset. Reference

[16] extracts features of each fault type sample to form a feature center for fault diagnosis. This general domain adaptation method has good fault diagnosis performance in solving unknown mechanical fault states. Reference

[17] extracts feature information through a Transformer network and uses threshold-based entropy maximum-minimum loss to further align shared health status samples, realizing accurate identification of bearing status in the general domain adaptation scenario.

[0008] However, the aforementioned fault diagnosis methods based on general domain adaptation overemphasize learning domain-invariant features and fail to fully utilize the target domain-specific data structures, which may result in the inability to learn useful features that distinguish unknown categories in the target domain from known source categories.

[0009] On the other hand, contrastive learning, as a representative method in the field of self-supervised learning, utilizes surrogate tasks to train models to distinguish between similar and different samples, thereby learning the general features of unlabeled data. Currently, numerous contrastive learning networks have been proposed, such as Simple Contrastive Learning of Representations (SimCLR).

[18] Momentum Contrast (MoCo)

[19] This involves constructing positive and negative samples to extract effective features. Meanwhile, self-predictive contrastive learning (Bootstrap Your Own Latent, BYOL) is another method.

[20] Only positive samples are needed for training, which reduces the need for datasets and thus improves the robustness of the model. Reference

[21] proposes a data augmentation method that combines truncation and downsampling, and uses the BYOL network to extract fault features from unlabeled vibration data to obtain a high diagnostic accuracy. Reference

[22] proposes a multi-objective contrastive learning method, which further improves the generalization of the model while realizing fault diagnosis.

[0010] Since contrastive learning does not consider the domain gap problem, directly using contrastive learning to extract rolling bearing features under different working conditions will result in the model failing to achieve feature alignment, which will easily lead to poor fault diagnosis performance. Furthermore, existing general domain adaptation methods rely too much on source domain label information, making it difficult to obtain discriminative features in the target domain. Summary of the Invention

[0011] To address the aforementioned problems, this invention proposes a method and system for diagnosing cross-domain faults in rolling bearings where the inter-domain data label relationship is unknown.

[0012] According to one aspect of the present invention, a method for cross-domain fault diagnosis of rolling bearings with unknown inter-domain data label relationships is proposed. The method includes the following steps:

[0013] 1) Data selection and preprocessing:

[0014] Vibration signals of rolling bearings with known labels under a certain working condition are obtained as source domain data, and vibration signals of rolling bearings under other working conditions without labels are obtained as target domain data. Short-time Fourier transform is performed on the source domain and target domain data to construct a two-dimensional image dataset, which is used as the model input.

[0015] 2) Comparison of general domain adaptation model construction:

[0016] The constructed contrastive general domain adaptation model consists of two modules: a contrastive learning module and a domain adaptation module. The contrastive learning module is used to enhance the model's feature extraction capability, maximize the similarity between samples with the same health status, and reduce the number of samples near the classification boundary. The domain adaptation module is used to solve the negative transfer when the data label relationship between domains is unknown, and to reduce the differences in data distribution and labels.

[0017] The target domain data after different data augmentations are input into the contrastive learning module to obtain the optimal encoder network. Then, the entropy separation strategy in the domain adaptation module is used to reject samples of unknown class in the target domain. The pseudo-labels of the unlabeled data in the target domain are obtained by using the source classifier. The transfer semantic augmentation method improved by the source class weighting mechanism is used to assign different class-level weights to each source class, thereby aligning the features of the shared classes in the source and target domains.

[0018] 3) Multi-state recognition:

[0019] During the training of the multi-state classification model, the LARS optimization algorithm is used to accelerate the model convergence speed and dynamically change the weight of the contrast loss. The model parameters are continuously updated through the target loss. Private health state samples are marked as unknown classes according to the threshold, while shared health state samples are correctly classified. After multiple iterations of optimization, a multi-state identification model for rolling bearings under different working conditions is established. The test data is input into the diagnostic model to obtain the final fault classification result.

[0020] Furthermore, the process of inputting target domain data with different data augmentations into the contrastive learning module to obtain the optimal encoder network includes:

[0021] For each image, two image enhancement methods are randomly selected, and the online network projection features z are calculated. θ ; Calculate the target network projection features z′ θ ; Calculate the loss L of the self-supervised representation learning method BYOL. CL After N iterations, the online network parameters are updated based on the gradient, the target network parameters are updated, and the optimal encoder network is obtained.

[0022] Furthermore, the process of rejecting samples of unknown class in the target domain using the entropy separation strategy in the domain adaptation module includes:

[0023] The distance between entropy H(y) and threshold ρ is defined as |H(y)-ρ|, where y is the classification output of the target sample; the entropy of unknown class samples is higher than ρ, and the entropy of shared class samples is lower than ρ. Using the confidence threshold parameter m, the entropy separation loss is calculated only for samples within the confidence interval.

[0024]

[0025]

[0026] Where B represents the batch size.

[0027] Furthermore, the process of assigning different class-level weights to each source category in the improved transfer semantic enhancement method through the source class weighting mechanism includes:

[0028] The weights of the source classes are estimated using the entropy values ​​of the target domain samples. The weights are designed as follows:

[0029]

[0030] Where G is the classifier, and Let n represent the weight of the c-th class and the target domain sample, respectively. t,c F is the number of target domain samples of category c; θ For encoder.

[0031] The weights are normalized using the minimum-maximum normalization method:

[0032]

[0033]

[0034] Where, ω c for The normalized value, where ψ is the scaling parameter. It is a small positive number;

[0035] Transfer semantic enhancement loss L CLS Parameters in Improved to:

[0036]

[0037] in, It is sample x si The probability of being predicted as category c; Represents category y si The difference in mean between regions of the sample; This represents the within-class covariance of samples predicted to be of class c; Represents category y si The weight matrix of the last fully connected layer;

[0038] To maximize the mutual information of the target domain data:

[0039]

[0040] Among them, C s Indicates the total number of source domain categories; n t Indicates the total number of samples in the target domain; This represents the probability that a sample in the target domain is predicted to be of class c.

[0041] Furthermore, during the training of the multi-state classification model, the weights of the contrastive loss are dynamically changed, and the dynamic tradeoff parameter β(x) is defined as:

[0042]

[0043] Where i is the current iteration number and N is the total number of iterations;

[0044] The final optimization objective of the model is as follows:

[0045] L = L CLS +α1L MI +α2L ES +β(x)L CL

[0046] Here, α1 and α2 are trade-off hyperparameters.

[0047] Furthermore, the online network projection feature z is calculated according to the following formula. θ :z θ ←g θ (F θ (t(x i ))).

[0048] Furthermore, the target network projection feature z′ is calculated according to the following formula. θ :z′ θ ←g′ θ (F′ θ (t′(x i ))).

[0049] Furthermore, the loss L of the self-supervised representation learning method BYOL is calculated according to the following formula. CL :

[0050]

[0051] Where, q θ (z θ1 ), z′ θ1 These are the output characteristics of the online network and the target network after exchanging views.

[0052] Furthermore, update the target network parameters according to the following formula: θ′←τθ′+(1-τ)θ.

[0053] According to another aspect of the present invention, a cross-domain fault diagnosis system for rolling bearings with unknown inter-domain data label relationships is proposed. The system includes:

[0054] The data preprocessing module is configured to acquire all state vibration signals of rolling bearings with known labels under a certain working condition as source domain data, and take the vibration signals of rolling bearings under other working conditions without labels as target domain data; and to construct a two-dimensional image dataset by performing short-time Fourier transform on the source domain and target domain data, which is used as model input.

[0055] The contrastive general domain adaptation model construction module is configured to input target domain data with different data augmentations into the contrastive learning module to obtain the optimal encoder network. Then, the entropy separation strategy in the domain adaptation module is used to reject samples of unknown classes in the target domain. The source classifier is used to obtain pseudo-labels for the unlabeled data in the target domain. A transfer semantic enhancement method improved by a source class weighting mechanism assigns different class-level weights to each source class, thereby aligning the features of shared classes between the source and target domains. The constructed contrastive general domain adaptation model consists of two modules: a contrastive learning module and a domain adaptation module. The contrastive learning module enhances the model's feature extraction capability, maximizes the similarity between samples with the same health status, and reduces the number of samples near the classification boundary. The domain adaptation module addresses negative transfer when the data label relationship between domains is unknown and reduces data distribution and label differences.

[0056] The state recognition module is configured to use the LARS optimization algorithm to accelerate model convergence during the training of the multi-state classification model and dynamically change the weight of the contrast loss; continuously update the model parameters through the target loss, mark private health state samples as unknown classes according to the threshold, and correctly classify shared health state samples; after multiple iterations of optimization, establish a multi-state recognition model for rolling bearings under different working conditions; input the test data into the diagnostic model to obtain the final fault classification result.

[0057] The beneficial technical effects of this invention are:

[0058] This invention employs contrastive learning to train target domain data, enabling the model to focus more on feature learning within the target domain. It also uses an entropy separation strategy to identify unknown health states within the target domain. Furthermore, a source class weighting mechanism is designed to improve the transfer semantic enhancement method, reducing differences in vibration data distribution and labeling. Finally, a Layer-wise Adaptive Rate Scaling (LARS) optimization algorithm is used to accelerate model convergence, achieving cross-domain fault diagnosis of rolling bearings with unknown inter-domain data label relationships. Attached Figure Description

[0059] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, which together with the detailed description below are incorporated in and form a part of this specification, and serve to further illustrate preferred embodiments of the invention and explain the principles and advantages of the invention.

[0060] Figure 1 This is a flowchart of a method for diagnosing cross-domain faults in rolling bearings where the inter-domain data label relationship is unknown, as described in this invention.

[0061] Figure 2 This is a schematic diagram of the BYOL structure.

[0062] Figure 3 This is a schematic diagram of the projector and predictor structure.

[0063] Figure 4 This is a schematic diagram of the model principle.

[0064] Figure 5 This is a schematic diagram of the entropy separation strategy process.

[0065] Figure 6 This is a schematic diagram of a bearing test bench.

[0066] Figure 7 This is an example diagram showing the experimental results comparing the method proposed in this invention with other methods.

[0067] Figure 8 These are example visualizations of t-SNE using different methods.

[0068] Figure 9 This is an example chart comparing the accuracy results before and after adding the source class weighting mechanism.

[0069] Figure 10 It is a confusion matrix that does not use contrastive learning.

[0070] Figure 11 It is a confusion matrix using contrastive learning.

[0071] Figure 12 This is an example image showing the comparison results of the optimization algorithms.

[0072] Figure 13 It is the source class weight of the method of this invention under task T5.

[0073] Figure 14 It is a sensitivity test of hyperparameters α1 and α2.

[0074] Figure 15 It is a sensitivity test of the hyperparameter m.

[0075] Figure 16 This is an example of the results of a cross-specification generalization comparison experiment. Detailed Implementation

[0076] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0077] This invention proposes a method for diagnosing cross-domain faults in rolling bearings where the inter-domain data label relationship is unknown. Figure 1 As shown, the specific steps of this method are as follows:

[0078] 1) Data selection and preprocessing.

[0079] Vibration signals of rolling bearings with known labels under a certain working condition are obtained as source domain data, and vibration signals of rolling bearings under other working conditions without labels are used as target domain data. Short-time Fourier transform is performed on the source domain and target domain data to construct a two-dimensional image dataset, which is used as the input of the model.

[0080] 2) Compare the construction of the general domain adaptation model.

[0081] The target domain data after different data augmentations are input into the contrastive learning module; then, the entropy separation strategy in the domain adaptation module is used to reject samples of unknown class in the target domain, and the pseudo-labels of the unlabeled data in the target domain are obtained by using the source classifier. The transfer semantic augmentation method improved by the proposed source class weighting mechanism is used to assign different class-level weights to each source class, thereby aligning the features of the shared classes in the source and target domains.

[0082] 3) Multi-state recognition.

[0083] During the training of the multi-state classification model, the LARS optimization algorithm is used to accelerate model convergence and dynamically change the weights of the contrastive loss. Model parameters are continuously updated using the target loss, and private health state samples are labeled as unknown classes based on a threshold, while shared health state samples are correctly classified. After multiple iterations and optimizations, a multi-state identification model for rolling bearings under different operating conditions is established. Test data is input into the diagnostic model to obtain the final fault classification results.

[0084] The embodiments of the present invention will be described in detail below.

[0085] 1. Comparative learning

[0086] Contrastive learning, a typical self-supervised learning method, aims to learn the general features of unlabeled data through a proxy task. Its core idea is to utilize data augmentation to achieve the proxy task, maximizing the feature similarity between augmented samples of the same image while minimizing the similarity between features from different images.

[23] This invention uses the BYOL algorithm from the contrastive learning method, which does not require negative sample pairs during training, thus improving model performance while avoiding excessive computation. The BYOL structure is as follows: Figure 2 As shown. BYOL utilizes an online network and a target network for model training. The online network consists of an encoder F... θ Projector g θ and predictor q θ The network consists of a ResNet-50 encoder and two-layer nonlinear multilayer perceptrons (MLPs) for the projector and predictor, with the structure as follows: Figure 3 As shown. The target network architecture and the one without a predictor q. θ The online networks are identical, but the network weight parameters are updated differently. The weights θ of the online network are updated using gradient descent, while the weights θ' of the target network are updated using the exponential moving average of the online network.

[0087] θ′←τθ′+(1-τ)θ(1)

[0088] Where τ is a hyperparameter, τ∈[0,1).

[0089] Using two different data augmentation methods t and t', different views v and v' of x are generated and input into encoder F respectively. θ and F θ The corresponding features are obtained. The BYOL network treats different augmented views of the same image as positive sample pairs, and the goal of the loss function is to minimize the similarity distance between positive sample pairs:

[0090]

[0091]

[0092]

[0093] Where, q θ (z θ ), z′ θ These are the output features of the online network and the target network, respectively. <·> represents the vector dot product operation, and ||·||2 represents L2 regularization.

[0094] Because the BYOL network branches are asymmetric, the views augmented with different data are swapped, allowing them to propagate forward along the other branch, and the loss L is calculated. θ′,θ The final loss of the BYOL network can be expressed as:

[0095]

[0096] Where, q θ (z θ1 ), z′ θ1 These are the output characteristics of the online network and the target network after exchanging views.

[0097] 2. Transfer semantic enhancement

[0098] Transfer semantic enhancement methods implicitly enhance source domain features by aligning them with the semantics of the target domain, thereby improving the generalization ability of classifiers.

[24] Due to domain offset, the mean features of source domain samples and target domain samples of the same category differ, reflecting the overall semantic deviation of that category. For category c, the average feature vectors of the source and target domains are calculated. and Pseudo-labels are used to address the issue of missing labels in the target domain data. This is achieved by analyzing the difference in means between domains. Reduce overall semantic bias and utilize intra-class covariance. Capture semantic changes in the target domain.

[0099] To train the model in an end-to-end manner, a memory bank M is used to cache all features from the source and target domains, thereby facilitating the calculation of Δμ. c and In iteration batch j, update the features and corresponding labels in M:

[0100]

[0101] Where i is batch B (j) The sample index within, These are the features and labels stored in M, respectively.

[0102] To perform effective cross-domain semantic enhancement, source domain features are arranged along a multivariate distribution. Various semantic transformations are performed randomly in the direction to generate enhanced features. While preserving the source domain labels, the features f of each source domain sample are... si Augment the model H times, and then train the model using cross-entropy loss on the augmented feature set:

[0103]

[0104] Where, n s C s These are the total number of samples in the source domain and the total number of categories, respectively. These are the weight matrix and bias vector of the last fully connected layer of the model, f. si It is the i-th sample x in the source domain si Its characteristics.

[0105] As the number of enhancements H approaches infinity, to avoid excessive computational costs, an infinite number of enhancement features are implicitly generated, and the upper limit of the loss is derived based on the law of large numbers:

[0106]

[0107]

[0108]

[0109] in, It is sample x si The probability of being predicted as category c.

[0110] 3. General Domain Adaptation

[0111] In domain adaptation, target domain samples often lack labels, making it impossible to determine the relationship between source and target domain data labels. To achieve cross-domain fault diagnosis when the data label relationship between domains is unknown, this invention uses universal domain adaptation to reduce distribution and label differences between data from different operating conditions.

[0112] In general domain adaptation, samples from the source and target domains located in the common label space are classified into shared classes, while samples from private categories in the target domain are classified into unknown classes. The task of general domain adaptation is to use labeled source domain data to align shared categories while identifying unknown categories.

[25] .set up They represent the components of n respectively. s The source domain consists of n labeled samples and n t The target domain consists of unlabeled samples. The source domain label space C s and target domain tag space C t The relationship is unknown, meaning the source domain data is collected from known mechanical failure modes, while the target domain data is sampled from some known and additional unknown failure modes. C = C s ∩C t It is a shared tag space for the source and target domains. These are the private label spaces of the source and target domains, respectively. Source domain data feature distribution P s and the target domain data feature distribution P t Different, i.e., P s ≠P t The commonality between the source and target domains is defined as the Jaccard distance in the label space.

[26] :

[0113]

[0114] Where ξ∈[0,1], when ξ=1, the general domain adaptation degenerates into the closed set domain adaptation; the smaller the value of ξ, the less class-level knowledge the source domain and the target domain share, and the more challenging the task is.

[0115] 4. Construction of the comparative general domain adaptation model in this invention

[0116] The general domain adaptation model is mainly divided into two modules: a contrastive learning module and a domain adaptation module. The contrastive learning module enhances the model's feature extraction capabilities, uncovers the unique structure of the target domain data, maximizes the similarity between samples with the same health status, and reduces the number of samples near the classification boundary. The domain adaptation module addresses the negative transfer problem when the data label relationship between domains is unknown, and reduces data distribution and label differences. Therefore, theoretically, the proposed model can identify unknown class samples while reducing domain differences. A schematic diagram of the model's principle is shown below. Figure 4 As shown.

[0117] 4.1 Comparative Learning Module

[0118] The contrastive learning module performs data augmentation operations such as random cropping, random grayscale adjustment, and color jittering on the input image to form positive sample pairs. It then utilizes the similarity between these positive sample pairs to learn more discriminative feature representations. The algorithm flow of the contrastive learning module is shown in Algorithm 1.

[0119]

[0120] 4.2 Domain Adaptation Module

[0121] 1) Entropy separation strategy

[0122] The main challenge of general domain adaptation is improving the performance of the source classifier using unlabeled target domain data. However, most general domain adaptation methods only attempt to increase the confidence of the shared class. To distance the unknown class in the target domain from the known class in the source domain, the confidence of the unknown class in the target domain needs to be reduced. Therefore, this invention uses an entropy separation strategy to reject samples of the unknown class in the target domain. Since the unknown class and the known class have no common attributes, the entropy output by the source classifier for the unknown class samples is greater than the entropy for the shared class samples. Based on this idea, the entropy separation strategy uses the entropy output by the source classifier to define the boundary between the unknown class and the shared class. The process of the entropy separation strategy is as follows: Figure 5 As shown.

[0123] The distance between entropy H(y) and threshold ρ is defined as |H(y)-ρ|, where y is the classification output of the target sample. The entropy of unknown class samples is higher than ρ, while the entropy of shared class samples is lower than ρ. This distance can be maximized to move H(y) away from ρ. However, in most cases, the threshold is ambiguous and may change due to domain shift. Therefore, using a confidence threshold parameter m, the entropy separation loss is calculated only for samples within the confidence interval.

[0124]

[0125]

[0126] Where B is the batch size. When |H(y)-ρ| is sufficiently large, the model is confident in its decisions regarding the unknown and shared classes.

[0127] 2) Improved transfer semantic enhancement

[0128] Entropy separation strategies can identify unknown and shared classes in the target domain, but still require aligning shared classes between the source and target domains. Contrastive learning can effectively extract low-level features from unlabeled data in the target domain, but the generalization ability of the added classifier is limited. Therefore, transfer semantic enhancement methods are used to align the feature distributions of data in the source and target domains, thereby improving the model's generalization. Since the label relationships between domain data are unknown, directly applying transfer semantic enhancement methods to general domain adaptation scenarios can lead to negative transfer. Therefore, this invention proposes a source class weighting mechanism to improve transfer semantic enhancement methods. The improved method weights the mean and covariance matrices, changing the direction of semantic transformation to... This promotes the alignment of sample distribution in the shared label space and suppresses negative migration caused by samples in the private label space.

[0129] When designing the source class weights, since the entropy value of the unknown class in the target domain is larger than that of the shared class, this invention uses the entropy value of the target domain samples to estimate the weights of the source classes. Specifically, to ensure that the shared class receives a higher weight and the private class receives a lower weight, the weights are designed as follows:

[0130]

[0131] Where G is the classifier. and Let n represent the weight of the c-th class and the target domain sample, respectively. t,c F is the number of target domain samples of category c. θ For encoder.

[0132] The weights are normalized using the minimum-maximum normalization method:

[0133]

[0134] Where, ω c for The normalized value, where ψ is the scaling parameter. It is a small positive number. Therefore, the transfer semantic enhancement loss L CLS Parameters in Improved to:

[0135]

[0136] in, It is sample x si The probability of being predicted as category c; Represents category y si The difference in mean between regions of the sample; This represents the within-class covariance of samples predicted to be of class c; Represents category y si The weight matrix of the last fully connected layer.

[0137] Mutual information measures the correlation between two random variables. A strong correlation between target domain features and predictions is beneficial for transfer semantic enhancement, making the extracted features more informative and containing more useful predictive semantics. Therefore, maximizing the mutual information of the target domain data is crucial.

[0138]

[0139] Among them, C s Indicates the total number of source domain categories; n t Indicates the total number of samples in the target domain; This represents the probability that a sample in the target domain is predicted to be of class c.

[0140] 3) Optimization Objectives

[0141] To avoid overemphasizing the bias introduced by contrastive learning during training, this invention dynamically changes the weights of the contrastive loss, allowing the model to focus on the classification task in the later stages of training, thereby improving the model's classification accuracy. The dynamic tradeoff parameter β(x) is defined as:

[0142]

[0143] Where i is the current iteration number and N is the total number of iterations.

[0144] In summary, the final optimization objective of the model is as shown in equation (19), and the objective function L is optimized by using the LARS optimization method with an adaptive learning rate, so that different layers of the model have different learning rates, reducing the problem of large gradient fluctuations in the model, thereby improving the optimization effect.

[0145] L = L CLS +α1L MI +α2L ES +β(x)L CL (19)

[0146] Here, α1 and α2 are trade-off hyperparameters.

[0147] 5. Experimental verification

[0148] The CWRU dataset used in the experiments of this invention comes from Case Western Reserve University in the United States.

[27] A schematic diagram of the test bench is shown below. Figure 6 As shown, it mainly consists of an electric motor, a torque sensor, a power meter, and an electronic controller. The motor drive end and the fan end are equipped with rolling bearings of specifications SKF6205 and SKF6203, respectively.

[0149] The experiment used a 16-channel data logger to collect bearing vibration signals at different speeds of 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm, with a sampling frequency of 12 kHz. The data included normal conditions and three fault conditions, with each fault condition further including three damage diameters, for a total of 10 categories. For ease of description of the dataset, the symbol N represents the normal condition. The descriptions of the nine fault conditions are shown in Table 1.

[0150] Table 1. Methods of representing experimental data

[0151]

[0152] The experimental dataset uses SKF6205 as an example, with 150 samples selected for each category, and each sample having a length of 1024 points. Specific details are shown in Table 2. The effectiveness of the proposed method was verified by setting up 12 different cross-rotation speed tasks. The experimental task settings are shown in Table 3; the source domain health status and target domain health status for different tasks were typically randomly selected.

[0153] Table 2 Bearing Experiment Data Set

[0154]

[0155] All experiments were conducted on a computer with an NVIDIA GeForce RTX 3060 GPU and 16GB of RAM. The LARS optimization method was used for training, with a learning rate of 0.01, a batch size of 16, weight parameters α1 and α2 both set to 0.1, a threshold parameter m set to 0.2, and 100 iterations. To reduce experimental randomness, each experiment was repeated five times, and the average value was used as the final result.

[0156] Table 3 Domain adaptation tasks at different speeds

[0157]

[0158] The method proposed in this invention is compared with the following four domain adaptation methods to verify the superiority of the proposed method:

[0159] 1) Domain Adversarial Training of Neural Networks (DANN)

[28] As a typical closed-domain adaptation method, it uses a domain discriminator for adversarial training to learn cross-domain invariant features.

[0160] 2) Domain Conditioned Adaptation Network (DCAN)

[29] As a closed-set domain adaptation method, it utilizes the domain-conditional channel attention mechanism to stimulate different convolutional channels, thereby enabling fault diagnosis when data distributions differ significantly.

[0161] 3) Universal Domain Adaptation (UAN)

[23] By using sample similarity for weighting, the source and target domains can share class adaptation.

[0162] 4) Universal Domain Adaptation using the One-vs-All Network (OVANet)

[30] By learning the minimum inter-class distance in the source domain, shared classes and unknown classes in the target domain can be identified.

[0163] Fault diagnosis results for different domain adaptation tasks are as follows: Figure 7 As shown in the comparison results, the proposed method has a good cross-domain diagnostic accuracy, with most results exceeding 92%.

[0164] The DANN and DCAN methods achieved good diagnostic performance in closed-set domain adaptation tasks T3, T6, and T9. However, these two methods cannot identify unknown class samples, and significant negative transfer occurs when private and shared classes exist simultaneously in the source and target domains. For example, in task T7, the diagnostic accuracy of the closed-set domain adaptation method is only about 50%, while the general domain adaptation methods UAN and OVANet achieve an accuracy of 78%, and the accuracy of the proposed method reaches 99.6%, showing a significant improvement over the aforementioned comparative methods. In the more challenging tasks T2, T11, and T12, where the commonality ξ is relatively small, the accuracy of the proposed method still reaches about 90%, an improvement of at least 20% compared to UAN and OVANet, indicating that the proposed method has more stable performance. In different domain adaptation tasks, the performance of the proposed method is superior to other methods, demonstrating its superiority in solving cross-domain problems.

[0165] To more intuitively evaluate the effectiveness of the proposed method, feature visualization will be implemented. Specifically, taking task T5 as an example, the t-distributions-to-chastic neighbor embedding (t-SNE) algorithm will be used to perform dimensionality reduction and visualization of the last layer features extracted by the model. The visualization results of different methods are shown below. Figure 8 As shown, different shapes represent different health conditions.

[0166] from Figure 8As shown in the five sub-figures, after visualization of the features extracted using the DANN and DCAN methods, features from different categories are mixed together, indicating that directly applying the closed-set domain adaptation method to the general domain adaptation scenario will result in a large number of misclassified samples. Furthermore, the feature distribution maps obtained using the UAN and OVANet methods show relatively clear class boundaries, but some overlapping features still exist. In contrast, after visualization of the deep features extracted by the proposed method, features of the same category cluster together, while features of different categories are far apart, and most unknown class samples cluster and separate from other classes. This further demonstrates that the proposed method can effectively distinguish features of different health states.

[0167] To verify the impact of each component of the proposed method on performance, comparative experiments were conducted before and after introducing the transfer semantic enhancement method, before and after improving the feature alignment method using the source class weighting mechanism, before and after improving the target domain network using contrastive learning, and before and after using the LARS optimization algorithm. To verify the effectiveness of using the transfer semantic enhancement method, due to space limitations, task T5 was used as an example. Feature alignment was performed using an MMD-based statistical difference measurement method and a traditional adversarial domain adaptation method based on a domain discriminator, and the diagnostic performance of the transfer semantic enhancement method used in this invention was compared. The results are shown in Table 4.

[0168] Table 4. Comparison of accuracy results for different feature alignment methods

[0169]

[0170] Comparing the accuracy rates of different feature alignment methods in Table 4, it is evident that the accuracy rate is lowest (only 31.4%) when no alignment method is used in cross-domain diagnosis, indicating a significant difference in feature distribution between the source and target domains. The accuracy rates using the transfer semantic enhancement method are 22.4% and 14.0% higher than the statistical difference measurement method and the adversarial domain adaptation method, respectively. This demonstrates that the transfer semantic enhancement method can better align features between the two domains, improving the accuracy of fault diagnosis.

[0171] To investigate the classification performance of the improved transfer semantic enhancement method by adding a source class weighting mechanism when the data label relationship between domains is unknown, a model without the source class weighting mechanism was set up for comparison. Specific experimental results are as follows: Figure 9 As shown.

[0172] Depend on Figure 9 It can be seen that the fault identification accuracy is significantly improved by using the source class weighting mechanism. In particular, in tasks T1 and T7, the accuracy improved by 12.3% and 10.2% respectively after adding the source class weighting mechanism. This proves that when there are significant differences in the feature distribution and label space between the source and target domains, the transfer semantic enhancement method improved by the source class weighting mechanism can maximize the alignment of features between the two domains in the shared label space.

[0173] To verify that contrastive learning can improve the model's classification performance, a confusion matrix visualization method is used to clearly reflect the accuracy of the proposed method in labeling private health states as unknown classes, as well as the correct identification of shared health states. The confusion matrix for task T5 is as follows: Figure 10 and Figure 11 As shown. Comparison Figure 10 and Figure 11 It can be seen that without contrastive learning, the diagnostic effect is poor and misclassification is quite obvious due to the insufficient exploitation of the target domain's unique data structure, resulting in a total of 23 misclassified samples. Using the method proposed in this invention, after introducing contrastive learning into the target domain for training, the number of misclassified samples is significantly reduced, with only four types of faults—IR14, OR14, B14, and unknown—exhibiting misclassification, totaling 13 misclassified samples. Experiments show that contrastive learning can enhance the ability to extract features of rolling bearing faults.

[0174] The experiment still uses task T5 as an example. During the model training process, the LARS optimization algorithm is used to improve the gradient convergence speed. The experimental results compared with the traditional optimization algorithm Adam are as follows: Figure 12 As shown.

[0175] Depend on Figure 12 It can be seen that when using the Adam optimization algorithm, the loss function exhibits severe fluctuations in the early stages of iteration, and the accuracy tends to stabilize around 80 iterations. When using the LARS optimization algorithm, the fluctuations in the loss function are not obvious in the early stages of iteration, and the accuracy remains stable around 30 iterations. The LARS optimization algorithm saves approximately 60% of the training time and can effectively accelerate the convergence speed.

[0176] Taking task T5 as an example, the source class weights learned by the model are studied, and the results are as follows: Figure 13 As shown in the diagram, observation reveals that the proposed method assigns greater weight to shared health states and less weight to private health states. This allows the method to focus on the adaptive results of the shared classes with higher weights while ignoring the private classes. Therefore, in cross-domain fault diagnosis, the proposed method can ensure alignment of shared classes and separation of private classes between the two domains.

[0177] Furthermore, the hyperparameter sensitivity of the proposed method was studied by changing the values ​​of the following parameters: 1) L MI The trade-off parameters α1; 2)L ES The trade-off parameter α2; 3)L ES The confidence threshold parameter m is used. To reduce the number of hyperparameters, α1 and α2 are specified to be equal in the experiment. The results are as follows: Figure 14 and Figure 15As shown, the diagnostic accuracy obtained within the given parameter range does not change significantly. Therefore, this indicates that the hyperparameters of the proposed method are insensitive within a reasonable range.

[0178] To further explore the generalization ability of the proposed method, experiments will be conducted on SKF6205, SKF6203, and the Jiangnan University Dataset (JNU).

[31] Fault diagnosis generalization experiments were conducted under different specifications.

[0179] The JNU sampling frequency is 50kHz, including four different health states. The experiment selected data from the four states at a rotation speed of 600 revolutions per minute, with 450 samples for each state.

[0180] To match the fault severity of the experimental datasets, four states of data from the SKF6205 and SKF6203 datasets were selected. The specific datasets and cross-specification task settings for the experiments are shown in Tables 5 and 6, respectively.

[0181] Table 5 Bearing Generalization Experiment Dataset

[0182]

[0183] Table 6 Domain Adaptation Tasks under Different Specifications

[0184]

[0185]

[0186] The experiment was conducted according to the task settings in Table 6, and the results are as follows: Figure 16 As shown. By Figure 16 It can be seen that in the six migration tasks, when classifying the fault states of bearings of different specifications, the fault diagnosis accuracy of the proposed method is higher than that of the other four domain adaptation methods, and the accuracy is improved by at least 13%. Experimental verification shows that the method proposed in this invention can not only solve the fault diagnosis problem at different speeds, but also solve the fault diagnosis problem of rolling bearings of different specifications, further proving that the method proposed in this invention has strong generalization ability.

[0187] 6. Conclusion

[0188] Based on the powerful feature extraction capabilities of contrastive learning, this invention proposes to fully exploit the unique structure of the target domain using contrastive learning. Through confusion matrix visualization, it is demonstrated that the improved target domain network enhances the feature extraction capability of rolling bearing faults. A source class weighting mechanism is proposed, and the transfer semantic enhancement method is improved to align the feature distribution of shared classes. Furthermore, an entropy separation strategy is introduced to reject samples from unknown classes. Experimental verification shows that the proposed improved transfer semantic enhancement method can improve the fault identification accuracy in general domain adaptation scenarios. Finally, a cross-domain fault diagnosis method for rolling bearings based on contrastive general domain adaptation is proposed, which can establish an effective fault diagnosis model even when the data label relationship between domains is unknown and labeled data in the target domain is lacking.

[0189] Experimental results show that the proposed method achieves high fault diagnosis accuracy, averaging 95.6%, even when there are significant differences in feature distribution and label space between the two domains. This represents an accuracy improvement of at least 8% compared to other domain adaptation methods compared in this invention. Furthermore, the proposed method employs the LARS optimization algorithm during training to accelerate model convergence, saving approximately 60% of training time. Generalization experiments demonstrate that the proposed method exhibits strong generalization ability across different transfer scenarios.

[0190] Another embodiment of the present invention proposes a cross-domain fault diagnosis system for rolling bearings with unknown inter-domain data label relationships. The system includes:

[0191] The data preprocessing module is configured to acquire all state vibration signals of rolling bearings with known labels under a certain working condition as source domain data, and take the vibration signals of rolling bearings under other working conditions without labels as target domain data; and to construct a two-dimensional image dataset by performing short-time Fourier transform on the source domain and target domain data, which is used as model input.

[0192] The contrastive general domain adaptation model construction module is configured to input target domain data with different data augmentations into the contrastive learning module to obtain the optimal encoder network. Then, the entropy separation strategy in the domain adaptation module is used to reject samples of unknown classes in the target domain. The source classifier is used to obtain pseudo-labels for the unlabeled data in the target domain. A transfer semantic enhancement method improved by a source class weighting mechanism assigns different class-level weights to each source class, thereby aligning the features of shared classes between the source and target domains. The constructed contrastive general domain adaptation model consists of two modules: a contrastive learning module and a domain adaptation module. The contrastive learning module enhances the model's feature extraction capability, maximizes the similarity between samples with the same health status, and reduces the number of samples near the classification boundary. The domain adaptation module addresses negative transfer when the data label relationship between domains is unknown and reduces data distribution and label differences.

[0193] The state recognition module is configured to use the LARS optimization algorithm to accelerate model convergence during the training of the multi-state classification model and dynamically change the weight of the contrast loss; continuously update the model parameters through the target loss, mark private health state samples as unknown classes according to the threshold, and correctly classify shared health state samples; after multiple iterations of optimization, establish a multi-state recognition model for rolling bearings under different working conditions; input the test data into the diagnostic model to obtain the final fault classification result.

[0194] The functionality of the rolling bearing cross-domain fault diagnosis system with unknown inter-domain data label relationships in this embodiment of the invention can be described by the aforementioned rolling bearing cross-domain fault diagnosis method with unknown inter-domain data label relationships. Therefore, for the parts not described in detail in the system embodiment, please refer to the above method embodiment, and they will not be repeated here.

[0195] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

[0196] The following documents are cited in this invention:

[0197] [1] Dong Shaojiang, Zhou Cunfang, Chen Lili, et al. Cross-domain fault diagnosis of bearings based on discriminative feature extraction and dual domain alignment [J / OL]. China Mechanical Engineering: 1-10 [2023-05-27]. http: / / kns.cnki.net / kcms / detail / 42.1294.TH.20230508.1159.004.html.

[0198] [2]ZHAO B,ZHANG X,Li H,et al.Intelligent fault diagnosis of rollingbearings based on normalized CNN considering data imbalance and variableworking conditions[J].Knowledge-Based Systems,2020,199:105971.

[0199] [3] SHI H, SHANG Y. Initial fault diagnosis of rolling bearing based on ondecond-order cyclic autocorrelation and DCAE combined with transfer learning[J]. IEEE Transactions on Instrumentation and Measurement, 2021,71:1-18.

[0200] [4] Jin Jiangtao, Xu Zifei, Li Chun, et al. Fault diagnosis of rolling bearings based on convolutional bidirectional long short-term memory network and chaos theory [J]. Vibration and Shock, 2022, 41(17):160-169.

[0201] [5] Liu, Haobo; Hao, Hongtao; Ding, Wenjie, et al. A method for fault diagnosis of rolling bearings based on IMCKD and MCCNN [J]. Vibration and Shock, 2022.

[0202] 41(07):241-249.

[0203] [6]WEN L, LI X, GAO L, et al. A new convolutional neural network-based data-driven fault diagnosis method[J].

[0204] IEEE Transactions on Industrial Electronics,2017,65(7):5990-5998.

[0205] [7] TONG J, TANG S, WU Y, et al. A fault diagnosis method of rollingbearing based on improved deep residual shrinkage networks [J]. Measurement, 2023, 206: 112282.

[0206] [8]WU Z,JIANG H,LIU S,et al.Conditional distribution-guided adversarial transfer learning network with multi-source domains for rollingbearing fault diagnosis[J].Advanced Engineering Informatics,2023,56:101993.

[0207] [9]ZHENG H,WANG R,YANG Y,et al.Cross-domain fault diagnosis using knowledge transfer strategy:A review[J].IEEEAccess,2019,7:129260-129290.

[0208]

[10] XIONG J, CUI S, TANG HA novel intelligent bearing fault diagnosis method based on signal process and multi-kerneljoint distribution adaptation[J]. Scientific Reports, 2023, 13(1):4535.

[0209]

[11] WU Z,JIANG H,LIU S,et al.A gaussian-guided adversarial adaptationtransfer network for rolling bearing fault diagnosis[J].Advanced EngineeringInformatics,2022,53:101651.

[0210]

[12] Wang Yujing, Lü Haiyan, Kang Shouqiang, et al. Fault diagnosis methods for different types of rolling bearings [J]. Proceedings of the CSEE, 2021, 41(01):267-276.

[0211]

[13] Chen Renxiang, Tang Linlin, Hu Xiaolin, et al. Fault diagnosis method for rolling bearings based on deep attention transfer learning at different speeds [J]. Vibration and Shock, 2022, 41(12):95-101.

[0212]

[14] Fan Cangning, Liu Peng, Xiao Ting, et al. A review of deep domain adaptation: general and complex cases [J]. Acta Automatica Sinica, 2021, 47(03): 515-548.

[0213]

[15] ZHANG W, LI

[0214]

[16] YAN Z,LIU G,WANG J,et al.A new universal domain adaptive method for diagnosing unknown bearing faults[J].Entropy,2021,23(8):1052.

[0215]

[17] ZHANG Y,REN Z,FENG K,et al.Universal source free domainadaptation method for cross-domain fault diagnosis ofmachines[J].MechanicalSystems and Signal Processing,2023,191:110159.

[0216]

[18] CHEN T, KORNBLITH S, NOROUZI M, et al. A simple framework for contrastive learning of visual representations[C]. International Conference on Machine Learning, Vienna, Austria, 2020: 1597-1607.

[0217]

[19] HE K,FAN H.Momentum contrast for unsupervised visualrepresentation learning[C].Proceedings of the IEEEConference onComputerVision andPattern Recognition,Seattle,WA,USA,2020:9729-9738.

[0218]

[20] GRILL J B,STRUB F,ALTCHE F,et al.Bootstrap your own latent-a newapproach to self-supervised learning[J].Advances in Neural InformationProcessing Systems,2020,33:21271-21284.

[0219]

[21] PENG T,SHEN C,SUN S,et al.Fault feature extractor based onbootstrap your own latent and data augmentation algorithm for unlabeledvibration signals[J].IEEE Transactions on Industrial Electronics,2021,69(9):9547-9555.

[0220]

[22] ZHANG W,CHEN D,XIAO Y,et al.Semi-supervised contrast learningbased on multi-scale attention and multi-target contrast learning forbearingfault diagnosis[J].IEEE Transactions on Industrial Informatics,2023,19:1-13.

[0221]

[23] KAN S,HE Z,CEN Y,et al.Contrastive bayesian analysis for deepmetric learning[J].IEEE Transactions on PatternAnalysis andMachineIntelligence,2022,45(6):7220-7238.

[0222]

[24] LI S,XIE M,GONG K,et al.Transferable semantic augmentation fordomain adaptation[C].Proceedings of the IEEE Conference on Computer Visionand Pattern Recognition,Nashville,TN,USA,2021:11516-11525.

[25] SU W,HAN Z,HER,et al.Neighborhood-based credibility anchor learning for universal domainadaptation[J].Pattern Recognition,2023,142:109686.

[0223]

[26] YOU K.,LONG M,CAO Z,et al.Universal domain adaptation[C].Proceedings ofthe IEEE Conference on Computer Vision and PatternRecognition,Long Beach,CA,USA,2019:2720–2729.

[0224]

[27] LOPARO,K A.Bearing data center[M / OL].Cleveland:Case WesternReserve University,2013.

[0225] http: / / csegroups.case.edu / bearingdatacenter / home.

[0226]

[28] GANIN Y,USTINOVA E,AJAKAN H,et al.Domain-adversarial trainingofneural networks[J].The Journal ofMachine Learning Research,2016,17(1):2096-2030.

[0227]

[29] LI S,LIU C,LIN Q,et al.Domain conditioned adaptation network[C].Proceedings ofthe AAAI Conference on Artificial Intelligence,New York,NewYork,USA,2020:11386-11393.

[0228]

[30] SAITO K,SAENTO K.Ovanet:One-vs-all network for universal domainadaptation[C].Proceedings of the IEEE Conference on Computer Vision andPattern Recognition,Nashville,TN,USA,2021:9000-9009.

[0229]

[31] LI K,PING X,WANG H,et al.Sequential fuzzy diagnosis method formotor roller bearing in variable operating conditions based on vibrationanalysis[J].Sensors,2013,13(6):8013-8041.

Claims

1. A cross-domain fault diagnosis method for rolling bearings with unknown inter-domain data label relationships, characterized in that, Comprising the following steps: 1) Data selection and preprocessing: Obtain all state vibration signals of rolling bearings under known labels of certain working conditions as source domain data, and obtain vibration signals of other working conditions of rolling bearings without labels as target domain data; perform short-time Fourier transform on the source domain and target domain data to construct a two-dimensional image data set as the model input; 2) Construction of a comparative universal domain adaptation model: The constructed comparative universal domain adaptation model comprises two modules: a comparative learning module and a domain adaptation module; The comparative learning module is used to enhance the feature extraction capability of the model, maximize the similarity between samples of the same health state, and reduce the number of samples near the classification boundary; the domain adaptation module is used to solve the negative transfer under the condition that the label relationship between the domains is unknown, and reduce the data distribution difference and label difference; The target domain data after different data enhancement is input into the comparative learning module to obtain an optimal encoder network; then, the entropy separation strategy in the domain adaptation module is used to reject unknown class samples of the target domain, the source classifier is used to obtain pseudo-labels of the target domain data without labels, and a transfer semantic enhancement method improved by a source class weighting mechanism is used to give different class-level weights to each source class, so as to align the features of the shared classes of the source domain and the target domain; 3) Multi-state recognition: In the training process of the multi-state classification model, the LARS optimization algorithm is used to speed up the convergence of the model, and the weight of the contrast loss is dynamically changed; the model parameters are constantly updated through the target loss, the private health state samples are marked as unknown classes according to the threshold, and the shared health state samples are correctly classified; after multiple iterations and optimization, a multi-state recognition model of rolling bearings under different working conditions is established; the test data is input into the diagnosis model to obtain the final fault classification result.

2. The method according to claim 1, wherein The process of inputting the target domain data after different data enhancement into the comparative learning module to obtain an optimal encoder network comprises: For each image, randomly select two augmented image methods, calculate the online network projection feature z θ ; Calculate the target network projection feature z' θ ; Calculate the loss L of the self-supervised representation learning method BYOL CL ; After N iterations, update the online network parameters according to the gradient, update the target network parameters, and obtain the optimal encoder network.

3. The method according to claim 2, wherein The process of rejecting unknown class samples of the target domain by using the entropy separation strategy in the domain adaptation module comprises: The distance between the entropy H(y) and the threshold ρ is defined as |H(y)-ρ|, wherein y is the classification output of the target sample; the entropy of the unknown class sample is higher than ρ, and the entropy of the shared class sample is lower than ρ; the confidence threshold parameter m is used to calculate the entropy separation loss only for the samples in the confidence interval: Wherein, B is the batch size.

4. The method according to claim 3, characterized in that, The process of giving different class-level weights to each source class by using the transfer semantic enhancement method improved by the source class weighting mechanism comprises: The weight of the source class is estimated by using the entropy value of the target domain sample, and the weight is designed as: where G is a classifier, and denote the weight of the c-th class and the target domain sample, respectively, t,c is the number of target domain samples of class c; F θ is an encoder; The weight is normalized by using the minimum-maximum normalization method: where ω c is the normalized value of the scaling parameter ψ, is a small positive number; Migration semantic enhancement loss L CLS parameters in The improvement is that: wherein, is a sample x si the probability of predicting a class c; denotes a class y si the domain-wise mean difference of a sample; denotes the within-class covariance of samples predicted as class c; denotes a class y si the weight matrix of the last fully connected layer of class y The mutual information of the target domain data is maximized as: where C s denotes the total number of source domain classes; n t denotes the total number of target domain samples; denotes the probability that a target domain sample is predicted to be class c.

5. The method according to claim 4, wherein In the training process of the multi-state classification model, the weight of the contrast loss is dynamically changed, and the dynamic weight balance parameter β(x) is defined as: Wherein, i is the current iteration number, and N is the total iteration number; The final optimization target of the model is as follows: L = L CLS + α1L MI + α2L ES + β(x)L CL Wherein, α1 and α2 are weight balance parameters.

6. The method according to claim 2, wherein The online network projection feature z is calculated according to the following formula θ : z θ ← g θ (F θ (t(x i ))).

7. The method according to claim 2, wherein The target network projected feature z' is calculated according to the following equation θ : z' = g'(F'(t'(x θ θ θ i )))​​​ 8. The method according to claim 2, wherein, The loss L of the self-supervised representation learning method BYOL is calculated according to the following formula CL : where q θ (z θ1 ), z′ θ1 are the output features of the online network and the target network after the view exchange.

9. The method according to claim 2, wherein, The target network parameters are updated according to the following formula: θ'←τθ'+ (1-τ)θ.

10. An inter-domain data label relationship unknown rolling bearing cross-domain fault diagnosis system, characterized in that, Comprise: The data preprocessing module is configured to obtain all state vibration signals of a rolling bearing with known labels of working conditions as source domain data and vibration signals of other working conditions without labels as target domain data; and perform short-time Fourier transform on the source domain and target domain data to construct a two-dimensional image data set as model input; The comparative universal domain adaptation model construction module is configured to input the target domain data after different data augmentations into a contrastive learning module to obtain an optimal encoder network; then reject unknown class samples of the target domain by using an entropy separation strategy in the domain adaptation module, obtain pseudo labels of the target domain unlabeled data by using a source classifier, and assign different class-level weights to each source class by using a transfer semantic enhancement method improved by a source class weighting mechanism, so as to align the features of shared classes of the source domain and the target domain; The constructed comparative universal domain adaptation model comprises two modules: a contrastive learning module and a domain adaptation module. The contrastive learning module is used to enhance the feature extraction capability of the model, maximize the similarity between samples of the same health state, and reduce the number of samples near the classification boundary; and the domain adaptation module is used to solve the negative transfer under the condition that the label relationship between domains is unknown and reduce the data distribution difference and label difference. The state recognition module is configured to use a LARS optimization algorithm to speed up the convergence speed of the model during the training process of the multi-state classification model and dynamically change the weight of the contrastive loss; constantly update the model parameters by using a target loss, mark private health state samples as unknown classes according to a threshold, and correctly classify shared health state samples; after multiple iterations of optimization, a multi-state recognition model of the rolling bearing under different working conditions is established; and test data is input into the diagnosis model to obtain a final fault classification result.

Images