Rolling bearing fault diagnosis method and system based on comparison decoupling single source domain generalization

Abstract

The invention discloses a rolling bearing fault diagnosis method and system based on comparison decoupling single-source domain generalization, relates to the field of rolling bearing fault diagnosis, and aims to solve the problem that in the prior art, a method for effectively solving the problem that the model generalization performance is reduced due to insufficient diversity of single-source data in bearing fault diagnosis does not exist. The method is technically characterized by comprising the following steps: step 1, adopting rolling bearing time domain data, and performing two data enhancement processing on the rolling bearing time domain data to obtain two groups of enhanced data; step 2, the model construction is based on a causal decoupling network and a comparative learning BYOL framework, the causal decoupling network is used for extracting causal features and non-causal features of data, and the rank of the causal features is constrained through adaptive threshold weighted nuclear norm regularization; the contrast learning BYOL framework is used for pulling in causal features extracted from the two groups of enhanced data to obtain causal domain invariant features of each sample; step 3, using a prototype Gaussian triple loss function to constrain causal domain invariant features, and improving intra-class compactness and inter-class separability of training samples; and step 4, finally inputting a target domain sample to the trained comparison decoupling network model to realize diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rolling bearing fault diagnosis, in particular to a rolling bearing fault diagnosis method and system based on contrast decoupling single-source domain generalization. BACKGROUND

[0002] As the core component in industrial rotating machinery, the stability of bearing performance is directly related to the operation safety of the whole system. Effective fault diagnosis of bearings is not only the key to prevent equipment failure and reduce downtime, but also an important means to ensure personnel safety and avoid major economic losses. With the rapid development of sensors and computer technology, deep learning (DL) based fault diagnosis methods are increasingly valued because they can automatically identify and extract key features [1] . However, the traditional DL strategy is based on the assumption that the training and testing data conform to the same distribution [2] . However, in actual scenarios, the working conditions of mechanical equipment are affected by various factors such as working load and speed, resulting in differences in data quality and data distribution, which affects the diagnostic performance of the above model.

[0003] To break the assumption of the same distribution, transfer learning is introduced into the cross-domain fault diagnosis task. Existing transfer research can be divided into two parts: domain adaptation (DA) [3] and domain generalization (DG) [4] . In the assumption of DA, the source domain and target domain data are used in the training stage, and by aligning the data distribution of the two domains, the fault diagnosis of the target domain is realized. Kang et al. proposed a contrast universal domain adaptation model for bearing fault diagnosis under the condition that the label space relationship between the target domain and the source domain is unknown [5] . DA-based methods rely on target domain data in model training process and diagnostic performance. In actual industrial scenarios, the operating conditions of mechanical equipment are unpredictable, and it is impossible to obtain target domain data before deploying the fault diagnosis model, which greatly limits the applicability of DA methods.

[0004] In view of the shortcomings of DA, DG-based fault diagnosis has emerged because it does not require target domain data to participate in training, and the model can complete training only on the source domain to diagnose the target domain data under unseen working conditions, so it is more efficient [6] . Most existing DG methods are based on training multiple source domains. Ragab et al. proposed a conditional contrast domain generalization method, which maximizes the mutual information of similar classes in multiple source domains and minimizes the mutual information between different classes, learns domain-independent feature representations, and realizes effective unknown field transfer [7]Ma et al. proposed a two-stage multi-source domain generalization diagnosis method inspired by causality, which used feature diversity activation and non-causal feature suppression to obtain more causal features while eliminating the interference of non-causal features [8] Wang et al. proposed a new domain generalization fault diagnosis method with multi-domain specific auxiliary classifiers, which can eliminate domain-specific features between source domains and extract cross-domain general diagnostic knowledge [9] However, in real industrial scenarios, it is very difficult to collect fault data under multiple working conditions. Usually, the available data comes from a single working condition, which severely limits the application of traditional domain generalization models in real scenarios. Compared with traditional multi-source domain generalization, single-source domain generalization (SDG) is more challenging in mechanical equipment diagnosis due to the single distribution of training samples.

[0005] Most existing SDG fault diagnosis researches use data augmentation techniques to expand the distribution range of the source domain by increasing the diversity of the source data, thereby strengthening the generalization of the model

[10] Han et al. proposed an internal and external hybrid domain generalization network (IEDG), which enhances single-source domain data by changing sample amplitude and adding Gaussian noise

[11] Zhao et al. proposed an SDG network guided by adversarial mutual information, which generates false target domains through a domain generation module and uses mutual information iterative minimax game to improve the resistance of domain transfer

[12] Tang et al. proposed a sequential embedding network based on histogram matching hybrid (HmmSeNet), which generates data with the same semantics but different distributions through histogram matching hybrid technology, thereby enhancing the single-source domain

[13] Wang et al. proposed a multi-scale style generation and adversarial contrastive network (MSG-ACN), which designed a multi-scale style generation strategy and an adversarial contrastive learning strategy, effectively enhancing the quality of sample generation in the single-source domain

[14] Li et al. proposed a dual adversarial and contrastive network (DACN), which realized pseudo-working condition feature generation and domain-invariant feature extraction through feature generation adversarial and cross-domain contrastive learning, and verified the diagnosis ability for unseen working conditions on industrial process data sets

[15] The above methods aim to generate as diverse source data distribution as possible to expand the single-source domain, thereby covering the unseen target domain. However, when the expanded source data distribution is significantly different from the target data distribution, the domain gap between the source domain and the target domain may still be significant, which will lead to a significant decline in the generalization performance of the model. Therefore, how to narrow the gap between the source domain and the target domain to improve the adaptability of the model in the unknown field has become a problem to be solved.

[0006] To improve the generalization ability of the model, the use of causal decoupling theory to explore cross-domain invariant features has become a key strategy to achieve domain generalization. Causal decoupling aims to separate the causal features directly related to the fault (such as the physical cause of mechanical failure) from the data, while stripping the non-causal features related to the domain (such as sensor noise and environmental differences)

[16] . DG fault diagnosis methods inspired by causal decoupling theory are gradually becoming a research hotspot, attracting more and more attention and exploration. Li et al. proposed a causal consistency network that mines invariant causal information in individualized machines through a collaborative approach to achieve knowledge generalization

[17] . Jia et al. proposed a causal domain disentanglement generalization network (CDDG) that disentangles data into causal and non-causal factors through a structural causal model, thereby achieving cross-domain diagnosis of bearing faults

[18] . Zhang et al. proposed a dual decoupling domain generalization method that separates domain-invariant and specific features through a shallow dual contrastive decoupling module, and separates relevant and irrelevant features through a deep adversarial mask decoupling module, combining KL divergence loss to constrain feature information integrity, achieving cross-domain fault diagnosis of rotating machinery

[19] . Guo et al. proposed a deep causal disentanglement network that decomposes fault data into causal and non-causal factors through a structural causal model, and combines causal aggregation loss and information entropy loss to enhance the model's ability to decouple cross-domain fault features

[20] . However, data may over-restrict non-causal but discriminative features during causal decoupling, leading to a decline in the classification ability of feature representation.

[0007] To address the problem of causal decoupling that may weaken discriminative features, contrastive learning can significantly improve the sufficiency of feature learning by enhancing the inter-class separability of features and strengthening the discriminative information between samples, and is widely used in the field of fault diagnosis

[21] . Liu et al. proposed a self-supervised contrastive learning model combining time and frequency domains for fault diagnosis

[22] . Research on fault diagnosis methods combining transfer learning and contrastive learning is also extensive and has achieved widespread application. For example, Pan et al. proposed a complementary weighted dual adversarial network combining supervised contrastive learning to solve the open-set domain adaptation fault diagnosis problem of bearings

[23] . However, when solving the domain generalization problem, contrastive learning relies on the statistical correlation of data (such as signal similarity), and may learn domain-specific false associations (such as signal amplitude differences caused by sensor location). When the target domain distribution changes, these surface features will fail, leading to a decline in the model's generalization ability.

[0008] Therefore, how to effectively solve the single-source domain generalization problem in bearing fault diagnosis is a pressing issue. SUMMARY

[0009] The technical problem solved by the present application is:

[0010] The prior art lacks an effective solution to the problem of low accuracy of rolling bearing fault diagnosis caused by single-source domain generalization in bearing fault diagnosis, therefore, the present application provides a rolling bearing fault diagnosis method and system based on contrast decoupling single-source domain generalization.

[0011] The technical solution adopted by the present application to solve the above technical problem is:

[0012] The present application provides a rolling bearing fault diagnosis method based on contrast decoupling single-source domain generalization, comprising the following steps:

[0013] Step one, adopt rolling bearing time domain data, and perform two kinds of data enhancement processing on it to obtain two groups of enhanced data;

[0014] Step two, the model is constructed based on a causal decoupling network and a contrast learning BYOL framework, the causal decoupling network is used to extract causal features and acausal features of data, and the rank of the causal features is constrained by self-adaptive threshold weighted kernel norm regularization, and the contrast learning BYOL framework is used to pull the causal features extracted from the two groups of enhanced data to obtain causal domain invariant features of each sample;

[0015] Step three, a prototype Gaussian triplet loss function is used to constrain the causal domain invariant features to improve the intra-class compactness and inter-class separability of the training samples;

[0016] Step four, finally, input the target domain sample to the trained contrast decoupling network model to realize diagnosis.

[0017] Further, step one includes the following process:

[0018] The rolling bearing time domain data is used to construct a time domain sample set with a single sample length of L points in a sliding sampling manner, and the enhanced sample set is obtained by adding noise, inverting the signal and window slicing data enhancement processing to each sample.

[0019] Further, step two separates causal features and acausal features, specifically:

[0020] First, the enhanced data sample is sent to the causal decoupling network, which introduces a structural causal model to describe the generation process of the signal:

[0021]

[0022] In the formula, represents the causal feature that has causal influence on and the fault label , and represents the acausal feature that influences , The structure function can be considered as unknown, is an unexplained noise variable.

[0023] Then, and respectively represent the causal feature matrix and the acausal feature matrix, and the dynamic weight redundancy reduction loss (HSIC) can be represented as:

[0024]

[0025] wherein, are the weights of each term, represents the HSIC operation, which is used to capture the nonlinear redundancy. The roles of the three terms in the above formula are to reduce the internal redundancy of causal features, to reduce the internal redundancy of acausal features, and to eliminate the nonlinear coupling between causal and acausal features. The calculation method of the weight can be represented as:

[0026]

[0027] wherein, is used to calculate the approximate rank of the feature matrix, is used to calculate the singular value of the matrix, , , , is a temperature parameter, taking a value of 0.5. The HSIC calculation method can be represented as:

[0028]

[0029] wherein, represents the L1 norm; and both adopt a radial basis function kernel, is a kernel bandwidth parameter, taking a value of 0.5, , represents a unit matrix, represents a full 1 vector.

[0030] After extracting features through the causal decoupling network, the constraint processing of HSIC can obtain causal features and acausal features.

[0031] Step two, constraint the rank of causal features, including the following processes:

[0032] The singular value decomposition is performed on the causal feature matrix to obtain each singular value , and the normalized energy ratio is calculated:

[0033]

[0034] Setting cumulative energy threshold (Default ), dynamically determine the number of principal components :

[0035]

[0036] Design a segmented weight function for principal components and noise:

[0037]

[0038] Among them, the principal component weight is small, reducing the suppression intensity of the previous singular value, retaining the main features; the noise weight is large, enhancing the punishment of the tail singular value, suppressing irrelevant components. Combine the weight with the singular value to define the adaptive threshold weighted kernel norm, which can be expressed as:

[0039]

[0040] Take the reciprocal of the kernel norm as the regularization term, so that minimizing the regularization term is to maximize the kernel norm. The adaptive kernel norm regularization loss is expressed as:

[0041]

[0042] In the formula, is an adjustment factor, denotes the number of samples;

[0043] Step two three, compare the decoupling model, including the following processes:

[0044] The training process of the contrast decoupling network is similar to the BYOL network. Two enhanced samples and of the original data are input into two branches respectively, and then the upper and lower branches learn from each other to minimize the distance between the output features of the two branches.

[0045]

[0046] Among them, denotes the predictor, denotes the projector, is the vector inner product, used to measure the cosine similarity of two normalized vectors, denotes the L2 norm.

[0047] Optimize the two branches of the network at the same time. In the asymmetric prediction link, the data of the upper and lower branches are exchanged, and the symmetric loss function is:

[0048]

[0049] The rolling bearing fault diagnosis method based on comparative decoupling and single-source domain generalization as described in claim 3 is characterized in that the prototype Gaussian triplet loss function constraining the causal domain invariant feature in step three specifically includes:

[0050]

[0051]

[0052] in, It is a sample. It is a feature extraction network Output the features of the sample. The feature prototype is represented and updated using the EMA method, where... Indicates the iteration number at the epoch-th iteration. The prototype of each category, Indicates the first Sample size for each category . It is the Gaussian kernel distance. It is right in front of you. It is a negative pair. This means that only the most difficult matching will result in a loss. This means that only the most difficult negative pair will cause a loss. This is called the margin, used to ensure that the distance difference between positive and negative samples has a certain interval. The Gaussian kernel distance function is defined as:

[0053]

[0054] in, express and The Euclidean distance between them. The standard deviation is expressed by the formula:

[0055]

[0056] A rolling bearing fault diagnosis system based on comparative decoupling and single-source domain generalization is provided, characterized in that the system has a program module corresponding to the steps of the method described in any one of claims 1 to 4, which executes the steps in the above-described comparative decoupling and single-source domain generalization rolling bearing fault diagnosis method during runtime.

[0057] A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program configured to implement, when invoked by a processor, the steps of the rolling bearing fault diagnosis method of comparative decoupling single-source domain generalization as described in any one of claims 1 to 4.

[0058] Compared with the prior art, the present application has the beneficial effects that:

[0059] To break through the difficulty of SDG fault diagnosis, the present application constructs a contrastive decoupling framework that combines the causal decoupling theory and the BYOL contrastive learning network. Specifically, the causal decoupling extracts causal feature representations that have a direct relationship with faults through feature separation technology, while eliminating non-causal features that are affected by domain conditions. Contrastive learning simulates potential domain differences by constructing data augmentation samples and uses a contrastive loss function to constrain the feature space at the feature decoupling and representation learning levels, enhancing the cross-domain invariance of causal features and forcing the differential expression of non-causal features. The combination of the two can provide dual protection at the feature decoupling and representation learning levels, enhancing the generalization performance of the model.

[0060] The present application proposes a contrastive learning architecture based on causal decoupling guidance, and constructs a contrastive decoupling single-source domain generalization network (CDSSDGN) model. The model constrains the contrastive feature generation process through a causal mechanism, aiming to solve the problem of fault diagnosis under new working conditions with single-source domain training. Specifically, first, the causal decoupling mechanism is combined with the BYOL (Bootstrap Your Own Latent) contrastive learning network to construct a contrastive decoupling framework. By embedding a dynamic weight redundancy reduction loss function based on the Hilbert-Schmidt independence criterion (HSIC), the correlation between causal and non-causal features is explicitly reduced, thereby improving feature separability. Then, an adaptive threshold weighted kernel norm regularization method is proposed, which can effectively improve the generalization of the model by constraining the rank of causal features. At the same time, to better achieve the intra-class compactness and inter-class separation of domain-invariant features, the feature prototype, Gaussian kernel function, and triplet loss function are combined, and the influence of outlier samples is considered, and a prototype Gaussian triplet loss function is proposed. Finally, the superiority of the proposed CDSSDGN model is verified through a large number of generalization diagnosis tasks on three bearing datasets.

[0061] The BYOL contrast learning network is combined with the causal decoupling theory to construct a contrast decoupling feature extraction network, and a Hilbert-Schmidt independence criterion is embedded to realize causal feature extraction; based on the low rank of the feature matrix in the kernel norm, by limiting the rank of the causal features learned from the training data, the model can be forced to focus on the causal features shared across domains, thereby strengthening the generalization performance of the model, and a prototype Gaussian triplet loss function is used to realize the intra-class compactness and inter-class separation of the causal domain invariant features, and the model still shows high performance when generalized to unknown working conditions. The application realizes accurate diagnosis of rolling bearing faults in rotating machinery, and lays a foundation for rolling bearing cross-condition generalization diagnosis and fault diagnosis of other components of rotating machinery. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 A single-source domain generalization schematic diagram in the embodiment of the application;

[0063] Figure 2 A bearing fault signal generation schematic diagram in the embodiment of the application;

[0064] Figure 3 A causal decoupling network in the embodiment of the application;

[0065] Figure 4 A BYOL architecture in the embodiment of the application;

[0066] Figure 5 A contrast decoupling single-source domain generalization framework diagram in the embodiment of the application;

[0067] Figure 6 A 1hp inner ring fault envelope spectrum diagram prototype Gaussian triplet loss schematic diagram in the embodiment of the application;

[0068] Figure 7 A CWRU bearing data acquisition platform in the embodiment of the application;

[0069] Figure 8 Generalization performance comparison of 8 methods on the CWRU dataset in the embodiment of the application;

[0070] Figure 9 Domain generalization result confusion matrix of 8 methods under task CT 1 in the embodiment of the application;

[0071] Figure 10 Feature visualization result of 8 methods under task CT 1 in the embodiment of the application;

[0072] Figure 11 A HUST bearing data acquisition platform in the embodiment of the application;

[0073] Figure 12Domain generalization results of 8 methods on HUST dataset in the embodiment of the application;

[0074] Figure 13 Confusion matrix of domain generalization results of 8 methods under task HT5 in the embodiment of the application;

[0075] Figure 14 Confusion matrix of domain generalization results of 8 methods under task HT5 in the embodiment of the application;

[0076] Figure 15 Domain generalization results of 8 methods under MT1 task in the embodiment of the application;

[0077] Figure 16 ROC curve of 8 methods under MT1 task in the embodiment of the application; DETAILED DESCRIPTION

[0078] In order to enable persons skilled in the art to better understand the present application scheme, in the following will be combined with the accompanying drawings Figures 1-16 Exemplary embodiments or examples of the present application are described. Obviously, the described embodiments or examples are only a part of the embodiments or examples of the present application, not all. Based on the embodiments or examples in the present application, all other embodiments or examples obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0079] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0080] In a typical embodiment, as shown in Figure 5 The present application provides a contrast decoupling single-source domain generalization rolling bearing fault diagnosis method, comprising the following steps:

[0081] Step one, time domain data of rolling bearing is adopted, and two data enhancement treatments are carried out, to obtain two groups of enhanced data;

[0082] Step two, the model construction is based on a causal decoupling network and a contrast learning BYOL framework, the causal decoupling network is used to extract causal features and acausal features of data, and the rank of causal features is constrained by self-adaptive threshold weighted kernel norm regularization, the contrast learning BYOL framework is used to pull the causal features extracted from the two groups of enhanced data, to obtain causal domain invariant features of each sample;

[0083] Step three, prototype Gaussian triplet loss function is used to constrain the causal domain invariant features, to improve the intra-class compactness and inter-class separability of training samples;

[0084] Step four, finally input the target domain sample to the trained contrast decoupling network model to realize diagnosis.

[0085] In an embodiment of the present application, step one includes the following process: as shown in the figure, time domain data of rolling bearings is used to construct a time domain sample set with a single sample length of 2048 points in a sliding sampling manner, and data enhancement processing such as adding noise, reversing signals and window slicing is performed on each sample to obtain an enhanced sample set. Figure 4

[0086] The process of extracting causal domain invariant features by using the contrast decoupling network in step two is as follows:

[0087] Step two one, separate causal features and acausal features, and the specific process includes:

[0088] Specific to the field of mechanical fault diagnosis, according to the literature

[25] , the bearing fault signal can be expressed by formula (1):

[0089]

[0090] In the formula, is the pulse caused by machine fault, is the interference factor such as electromagnetic interference, noise, etc. and are the transmission path effects of and respectively. It is assumed that does not contain interference information, and does not contain fault-related information. This assumption shows that the causal feature related to the fault can be separated from the acausal feature which is not related to the fault.

[0091] Figure 2 is the causal structure diagram of the bearing fault signal generation, where represents the fault mechanism, represents the machine, represents the signal, represents the label. As can be seen from formula (1), the causal relationship of the fault in the signal is concentrated on , but the signal itself is a highly coupled term of the fault and the machine . Therefore, the signal can be decomposed into causal features and acausal features . The causal structure diagram clearly shows the causal relationship between various factors, which helps to reveal the direct and indirect influencing factors of the fault signal. The SCM is introduced to describe the bearing fault signal generation process, which can be expressed by formula (2):

[0092] ​

[0093]

[0094] where, is the label, can be regarded as an unknown structure function, is the noise variable.

[0095] The training process of the contrast decoupling network is similar to that of the BYOL network. Two augmented samples of the original data and are input into two branches, respectively, and then the upper and lower branches learn from each other to minimize the distance between the output features of the two branches

[26] . Subsequently, the upper branch obtains the causal feature and the acausal feature as and , and the lower branch obtains and .

[0096] To reduce the redundant information in the causal and acausal features, a dynamic weight redundancy reduction loss based on HSIC is proposed by referring to the idea of the redundancy reduction loss function in document

[27] . Let and represent the causal feature matrix and the acausal feature matrix, respectively. The dynamic weight redundancy reduction loss can be expressed as:

[0097]

[0098] where, are the weights of each term, denotes the HSIC operation, which is used to capture the nonlinear redundancy. The roles of the three terms in the above formula are to reduce the internal redundancy of the causal feature, reduce the internal redundancy of the acausal feature, and eliminate the nonlinear coupling between the causal and acausal features. The calculation method of the weight can be expressed as:

[0099]

[0100] where, is used to calculate the approximate rank of the feature matrix, is used to calculate the singular value of the matrix, , , , is the temperature parameter, which takes the value 0.5. The HSIC calculation method can be expressed as:

[0101]

[0102] where, denotes the L1 norm; and both employ the radial basis function kernel, is the kernel bandwidth parameter, taking the value 0.5, , denotes the identity matrix, denotes the all-ones vector.

[0103] After extracting features through the causal decoupling network, the constraint processing of HSIC can obtain causal features and acausal features.

[0104] Step two, constraint the rank of causal features, includes the following process:

[0105] As the core factor of cross-domain sharing, the feature matrix of causal features has low rank in the ideal case. However, the noise contained in the causal features will introduce high rank interference. Therefore, by limiting the rank of the causal features learned from the training data, the model can be forced to focus on the causal features shared across domains, thereby strengthening the generalization performance of the model.

[0106] Mathematically, the calculation of the rank of a matrix is usually replaced by the kernel norm. The kernel norm of a matrix is defined as:

[0107]

[0108] where, , and denote the number of rows and columns of the matrix , respectively, denotes the th singular value of .

[0109] Large singular values reflect the main information of the feature matrix, while small singular values reflect the detailed information. Therefore, when maximizing , it is attempted to moderately reduce the weight of large singular values and increase the weight of small singular values. The basic principle of weight adjustment is: the larger the singular value, the smaller the coefficient. Therefore, according to the singular value itself, an effective weighting method is proposed:

[0110]

[0111] where, is the adjustment factor, is a non-negative integer. In this formula, will be regarded as the weight of . We can see that, The larger, the smaller the weight. Traditional weighted kernel norm regularization suppresses singular values by fixing the exponential decay, which may over-weak the principal components. To this end, an adaptive weighting strategy based on dynamic threshold division is proposed, aiming to distinguish principal components from noise and impose differentiated weight constraints respectively.

[0112] Singular value decomposition is performed on the matrix to obtain each singular value , and the normalized energy ratio is calculated:

[0113]

[0114] Set the cumulative energy threshold (default ), dynamically determine the number of principal components :

[0115]

[0116] Design a segmented weight function for principal components and noise:

[0117]

[0118] Where the principal component weight is small, reducing the suppression intensity of the previous singular value, preserving the main features; the noise weight is large, enhancing the punishment of the tail singular value, suppressing irrelevant components. Combine the weight with the singular value to define the adaptive threshold weighted kernel norm, which can be expressed as:

[0119]

[0120] Take the reciprocal of the kernel norm as the regularization term, so that minimizing the regularization term is to maximize the kernel norm. The adaptive regularization kernel norm can be expressed as:

[0121]

[0122] In the formula, is the adjustment factor, denotes the number of samples;

[0123] Step two Three, compare the decoupling model, including the following processes:

[0124] The training process of the contrast decoupling network is similar to the BYOL network, and the two enhanced samples and The upper and lower branches learn from each other to minimize the distance between the two branches' output features.

[0125]

[0126] where, denotes the predictor, denotes the projector, is the vector inner product, which is used to measure the cosine similarity of two normalized vectors, denotes the L2 norm.

[0127] To optimize the two branches of the network simultaneously, the data of the upper and lower branches are exchanged in the asymmetric prediction link, and the symmetric loss function is:

[0128]

[0129] The prototype Gaussian triplet loss function in step three constrains the causal domain-invariant features, which is specifically:

[0130] In order to better learn the generalization ability of the feature prototype enhancement model to unknown working conditions, it is not enough to use only the Euclidean distance to describe the complex relationship between samples. This paper combines the metric learning method

[30] and the kernel method

[31] to propose a prototype Gaussian triplet loss function (Proto Gaussian Triplet Loss, PGTL). Consider the pairing of class prototypes and each sample in each training batch, and map the samples to a higher dimensional space to obtain better classification results. The principle diagram of the PGTL function is shown in Figure 6 .

[0131] The definition of the triplet loss is:

[0132]

[0133]

[0134] where, is the sample, is the feature extraction network outputs the feature of the sample. denotes the feature prototype, which is updated by the EMA method in formula (17), where denotes the prototype of the i-th class at the epoch-th iteration, denotes the sample quantity of the i-th class, .​​ is the Gaussian kernel distance. is positive pair, is negative pair. indicates that only the most difficult positive pair will cause loss, indicates that only the most difficult negative pair will cause loss. is called margin, which is used to ensure that the distance difference between positive and negative samples has a certain interval. The Gaussian kernel distance function is defined as:

[0135]

[0136] where, indicates and the Euclidean distance between them. is the standard deviation, and its formula is:

[0137]

[0138] In this paper, five different values of standard deviation (i.e. ) are used, and the function value ranges from 0 to . The more similar the two samples are, the closer the function value is to 0. According to the proposed PGTL, for each sample , all samples from the same class are constructed as positive pairs with the class prototype, while all samples from different classes are used as candidate samples to construct negative pairs. Only the most difficult candidate positive pair will be used to calculate the loss.

[0139] The present application is based on the bearing fault diagnosis method of contrast decoupling single source domain generalization, as shown in Figure 5 , the core component of the framework is the deep integration of causal decoupling mechanism and BYOL contrast learning framework, which realizes cross-domain invariant feature extraction through the construction of double-branch decoupling contrast structure. In addition, the adaptive weighted kernel norm regularization strategy strengthens the generalization of the model by constraining the low rank of the causal feature matrix; at the same time, the prototype Gaussian triplet loss function is designed to learn the intra-class compactness and inter-class separability feature distribution of each fault class, so as to enhance the generalization ability of the model under unknown working conditions. The specific process is as follows:

[0140] The bearing fault diagnosis method (algorithm) based on contrast decoupling single source domain generalization proposed in the present application is the underlying technical core of the present application, and various products can be derived based on the algorithm.

[0141] The method based on the present application develops a rolling bearing fault diagnosis system based on contrast decoupling single-source domain generalization by using a programming language. The system has program modules corresponding to the steps of the above technical solutions, and executes the steps of the above rolling bearing fault diagnosis method based on contrast decoupling single-source domain generalization when running.

[0142] The computer program of the developed system (software) is stored on a computer readable storage medium, and the computer program is configured to realize the steps of the above rolling bearing fault diagnosis method based on contrast decoupling single-source domain generalization when called by a processor. That is, the present application is materialized on a carrier to become a computer program product.

[0143] Various embodiments of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0144] Computational programs (also referred to as programs, software, software applications, or code) in the present application include machine instructions for programmable processors, and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0145] The advantages of the present application will be described below in combination with specific examples and comparative examples.

[0146] To verify the performance of the CDSSDGN method described herein, experiments were conducted on three data sets, (1) the bearing data set of different rotating speeds of CWRU (Case Western Reserve University)

[32] ; (2) the bearing data set of different loads collected by HUST (Hanoi University of Science and Technology)

[33] ;

[0147] To ensure fairness, all experiments were run on the same equipment. The equipment configuration parameters are as follows: CPU: Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz; Graphics card: NVIDIA GeForce RTX 4090; running on Windows system.

[0148] Example 1: Dataset of rolling bearings at different speeds

[0149] CWRU bearing data acquisition platform, such as Figure 7 As shown, an accelerometer with a sampling frequency of 12kHz was installed at the drive end to collect bearing vibration signals. The dataset contains four rotational speeds: 1730 r / min, 1750 r / min, 1772 r / min, and 1797 r / min. This experiment aims to test the single-source domain generalization diagnostic performance of the proposed method at different rotational speeds.

[0150] The datasets collected at each rotational speed are considered as one domain, which can be divided into four domains: A, B, C, and D. Each domain contains 600 samples for each health state, with each sample containing 2048 data points, of which 80% are training samples. Detailed data information for each domain is shown in Table 1.

[0151] Table 1. Detailed information on single-source domain data in the CWRU dataset.

[0152]

[0153] Twelve domain generalization tasks were set up for the CWRU dataset, as shown in Table 2.

[0154] Table 2 SDG Fault Diagnosis Task Settings Based on CWRU Dataset

[0155]

[0156] Taking task CT1 in Table 2 as an example, dataset A is used as the source domain and dataset B is used as the target domain to test the generalization ability of the proposed method and the comparison method at different rotation speeds.

[0157] Domain generalization diagnostic results analysis:

[0158] To verify the effectiveness of the proposed CDSSDGN method, a single-source domain generalization experiment was conducted under the experimental settings described in Section 4.2.1. The proposed method was also compared with the method described in Section 4.1. The SDG experimental results of the proposed method and the comparison method on the CWRU dataset are shown below.

[0159] Table 4 Generalization results of 8 methods on the CWRU dataset

[0160]

[0161] According to the experimental results of Table 4 and Figure 8 Among all the methods, the CDSSDGN method proposed in this paper can achieve the highest domain generalization diagnostic accuracy, with an average accuracy of 97.99% and an average standard deviation of only 0.04%. This fully demonstrates the strong generalization ability of the CDSSDGN method, and also has strong stability.

[0162] The average accuracy of the CNN method in the 12 tasks is 61.34%, except for the accuracy of 84.06% in the CT9 task, the accuracy in the remaining tasks is less than 80%. The IEDG and HmmSeNet methods use data augmentation strategy to generate diversified source domain data, with an average accuracy of 70.76% and 88.15%, respectively. Under the same experimental conditions, the MSG-CAN and UCL-SDG methods using data augmentation and contrastive learning combination have an average accuracy of 94.19% and 90.42% respectively in the 12 single-source domain generalization tasks, which is better than the above three methods. This shows that the contrastive learning strategy can strengthen the model to extract domain-invariant features, and further improve the generalization performance of the model based on data augmentation. The average accuracy of the DACN method is 86.61%, which is lower than the average accuracy of the HmmSeNet method. This shows that DACN as a single-source domain generalization diagnostic method in process industry systems has certain limitations when applied to bearing diagnosis. CDDG extracts domain-invariant features through causal learning mechanism, with an average accuracy of 85.03%, which is lower than that of HmmSeNet and other methods, but the diagnostic performance is greatly improved compared to IEDG. Its average standard deviation is 0.71, and the model stability is excellent among existing methods, only slightly lower than the CDSSDGN method proposed in this paper. This shows that the causal decoupling mechanism can effectively learn generalization representation, while maintaining high stability while improving the generalization ability of the model.

[0163] The CDSSDGN method proposed in this paper combines the contrastive learning strategy and the causal decoupling mechanism, further designs the HSIC dynamic weight redundancy reduction loss, and effectively combines the ATWNNR and PGTL methods, so that it can fully extract cross-domain invariant representation. Among all the methods participating in the comparison, not only the highest diagnostic accuracy can be achieved, but also the highest model stability.

[0164] Diagnostic result visualization:

[0165] From Figure 9 It can be seen from (a) that the CNN method has confusion in identifying multiple categories, and only categories 4 and 6 are correctly identified. Figure 9In (b), the IEDG method has a great improvement compared to the CNN method overall, but almost all of the class 1 is misclassified, and there is a large overlap in identifying class 7, class 8 and class 9. Figure 9 As can be seen in (c), the CDDG improves the identification accuracy of other health states, but nearly 88% of the samples are misclassified as class 7 when classifying class 9. Figure 9 In (d), the HmmSeNet method also misclassifies most of the class 9 samples as class 7. The UCL-SDG and DACN methods misidentify class 8. The MSG-ACN method only misidentifies 6 B21 samples as O14 class, and the rest are all correctly classified. The proposed CDSSDGN method can accurately identify all health states, which has obvious advantages compared to other methods.

[0166] Also taking task CT1 as an example, the distribution of the features learned by each method is observed using t-SNE visualization technology. As shown in Figure 10 S and T respectively refer to the source domain and the target domain. Figure 10 In (a), the features learned by the CNN are not well clustered, and the target domain features cannot coincide with the source domain features, which is consistent with the Figure 9 poor classification results in (a). Figure 10 In (b), the confusion phenomenon of the IEDG method is better than the CNN, but the source domain features and the target domain features still cannot coincide well. The CDDG method can coincide well overall, but there is a significant overlap between B07 (class 7) and B21 (class 9), which confirms the results in Figure 9 (c). The target domain features of most classes of the HmmSeNet method cannot overlap with the source domain features, which shows that the sample features after using the data enhancement method deviate from the source domain feature distribution. The visualization results of the MSG-ACN method show that it can better achieve feature clustering, and only the target domain features of the B21 class cannot coincide with the source domain features. The results of the UCL-SDG and DACN methods show that the source domain and target domain features of most classes can coincide, but there is a feature overlap phenomenon, for example, the target domain sample B14 (class 8) of the UCL-SDG method is mixed with the I14 (class 2) and O14 (class 5) samples, which is consistent with the Figure 9 classification results in (f). As shown in Figure 10 (h), the proposed CDSSDGN can effectively learn domain-invariant representations from a single domain, can completely separate different health states, and the source domain and target domain features of the same class can coincide well. Compared with other methods, it shows better inter-class separability and intra-class aggregation.

[0167] Example 2: Rolling bearing dataset under different loads

[0168] The HUST dataset comes from the Hanoi University of Science and Technology, and the experimental platform is shown in Figure 11 The bearing model used in the experiment is SKF6205, and vibration data under three load conditions of 0W, 200W and 400W are collected respectively. All data are collected at a sampling frequency of 51.2kHz. This experiment aims to test the diagnostic performance of the proposed method under different loads.

[0169] The detailed information of the HUST dataset is shown in Table 5. The data collected under each load constitutes a domain, and the number of samples for each healthy state is 600, and each sample contains 2048 data points. Six domain generalization fault diagnosis tasks are set for the HUST dataset, as shown in Table 6.

[0170] Table 5 Details of single-source domain data of HUST dataset

[0171]

[0172] Table 6 Domain generalization fault diagnosis task settings based on HUST dataset

[0173]

[0174] Domain generalization diagnosis result analysis:

[0175] According to the above experimental settings, the single-source domain generalization experimental results of the method CDSSDGN proposed in this paper and the comparison methods on the HUST dataset are shown in Figure 12 The upper part of the box in the box plot is the average accuracy of each method, and the lower part is the sample standard deviation.

[0176] According to the experimental results in Figure 12 Among all the methods, the CDSSDGN method proposed in this paper can achieve an average accuracy of 98.81% in all tasks, and the average standard deviation is only 0.15%. This fully shows the strong generalization ability of the CDSSDGN method, and also has strong stability.

[0177] CNN performs poorly in some tasks (e.g., HT5), and although it has a high average accuracy in the HT4 task, it is unstable with large fluctuations. The IEDG and CDDG methods have some competitiveness in the HT4 and HT6 tasks, but their effectiveness is poor in other tasks. However, the average standard deviation of the CDDG method across all tasks is 0.74, which is only second to HmmSeNet and the proposed CDSSDGN method in terms of stability. Combined with the results of the causal decoupling mechanism in Case 1, it is shown that the causal mechanism can enhance the stability of model generalization. HmmSeNet, MSG-CAN, and DACN methods have good comprehensive performance in the HTI, HT2, HT4, and HT6 tasks, and even better than the proposed CDSSDGN method. However, the standard deviations of the MSG-CAN method in the HT3 and HT5 tasks are 4.51 and 4.04, respectively, and the diagnostic accuracy of the HmmSeNet method in the HT5 task is only 90.97%. This shows that HmmSeNet and MSG-CAN methods cannot fully extract domain-invariant features when training models on fault data with large loads, resulting in poor diagnostic performance in small load conditions.

[0178] Diagnosis result visualization:

[0179] To observe the details of the diagnosis results, we take task HT5 as an example and give the confusion matrix of the diagnosis results of the comparative methods and the proposed method, as shown in Figure 13 (a-h). Figure 13 In the above table, except for the UCL-SDG method, other methods have the phenomenon of misclassifying the healthy state B as the healthy state IO. However, in Figure 13 (f), the UCL-SDG method is prone to misclassifying the healthy state IO as the healthy states O and OB. Although CDSSDGN also causes confusion in identifying the healthy states B and IB, it improves the classification accuracy of the B category by an average of 46.97%, and has obvious advantages over other methods in identifying other healthy states and overall diagnostic accuracy.

[0180] Taking task HT5 as an example, we use t-SNE technology to observe the distribution of the features learned by each method. As shown in Figure 14 , T and S represent the target domain and the source domain, respectively. From Figure 14 (a), we can see that the features learned by CNN are not well clustered. Except for the UCL-SDG method, other methods have the phenomenon of confusion in identifying the healthy states B and IO. In addition, IEDG also has confusion in identifying the healthy states B and IB, which is consistent with Figure 13conclusions in the previous section. The features obtained by the HmmSeNet method are not aliased, but the source domain features for healthy states N, B, and I cannot overlap with the target domain features. This indicates that the samples generated by the HmmSeNet method deviate from the distribution of the original data. The DACN method results also show that it misclassifies healthy state I B as healthy state O. The proposed CDSSDGN can effectively learn domain-invariant representations from a single source domain, and it exhibits better inter-class separability and intra-class cohesion compared to other methods.

[0181] Example 3: Rolling bearing dataset under different loads

[0182] The MUET dataset includes three health conditions: healthy, inner race fault with different damage levels, and outer race fault. The vibration data of the bearing running under 100W, 200W, and 300W conditions were collected at a sampling rate of 10 kHz, respectively. In this experiment, the outer race data with different damage levels under three loads were used for single-source domain generalization experiments. The detailed information of the MUET data is shown in Table 7. Six domain generalization fault diagnosis tasks were set for the MUET rolling bearing dataset, as shown in Table 8.

[0183] Table 7 Detailed information of single-source domain data for MUET outer race fault

[0184]

[0185] Table 8 Domain generalization fault diagnosis task settings based on the MUET dataset

[0186]

[0187] Analysis of domain generalization diagnosis results

[0188] According to the above experimental settings, the single-source domain generalization experimental results of the proposed CDSSDGN method and the comparison methods on the MUET dataset are as follows.

[0189] Table 9 Diagnostic accuracy of eight methods on the MUET dataset

[0190]

[0191] From Table 9 and Figure 15It can be seen that the average diagnostic accuracy of the proposed CDSSDGN method is 83.33% with a standard deviation of 0.25 in all tasks, and it leads in 6 of the 8 tasks, especially in the most challenging MT5 task, with an accuracy of 69.42%, which is significantly higher than the HmmSeNet method by 13.67 percentage points. This shows that the proposed CDSSDGN method still has strong generalization and stability in the face of extremely complex SDG tasks with the same fault location and multiple fault levels. In contrast, the overall performance of CCN and IEDG methods is poor, with an average accuracy of less than 65%, and only IEDG performs well in the MT3 task. While DCAN and UCL-SDG methods perform well in some tasks, such as DCAN with an accuracy of 94.13% in the MT1 task, their average standard deviations in the 6 tasks are 2.32 and 2.21, respectively, indicating poor model stability. From the radar chart, it can be seen that the area enclosed by CDSSDGN contains all other methods except in the MT3 task. This reflects its high level of comprehensive performance in multiple tasks, indicating that the CDSSDGN method has strong comprehensive performance.

[0192] Diagnosis result visualization:

[0193] Figure 16 (a-h) show the ROC curves of the 8 methods in the MT1 task, which are used to evaluate the single-source domain generalization diagnosis results of each method. The area under the curve (AUC) value of each class quantifies the specific diagnostic performance. It is clear that the proposed CDSSDGN method shows the best diagnostic performance. All classes achieve a perfect AUC score of 1.00, indicating a sensitivity and specificity of 100%. In addition, the micro-average and macro-average ROC curves achieve the best AUC score of 1.00. These results show that the proposed method can distinguish between different classes in this task. The misdiagnosis in the comparative methods mainly occurs in the data of three fault levels: class 1 (D09), class 3 (D13), and class 5 (D17). The possible reason is that the features of these three fault levels are not prominent, which are easily confused and lead to incorrect classification.

[0194] Ablation experiment

[0195] To verify the effectiveness of the innovative part in the proposed method, ablation experiments are designed. Compared with the decoupling framework, the HSIC dynamic weight redundancy reduction loss is used to extract the cross-domain invariant causal representation in the source domain data; the adaptive threshold weighted kernel norm regularization (ATWNNR) forces the model to focus on the cross-domain shared causal features by constraining the rank of the causal features; the prototype Gaussian triplet loss (PGTL) function constrains the similarity between the intra-class feature distribution and the prototype center, and enhances the model's generalization ability to unknown working conditions. Table 10 and Table 11 show the diagnostic performance of the ablation experiments on the HUST dataset and the MUET dataset.

[0196] Table 10. Comparison of results of ablation methods on the HUST dataset

[0197]

[0198] Table 11. Comparison of results of ablation methods on the MUET dataset

[0199]

[0200] From Table 10 and Table 11, it can be seen that on the two datasets, the average accuracy of M1 is improved by 1.24% and 11.02% respectively compared with the baseline method. In each task, the diagnostic performance of the decoupling framework combined with the HSIC dynamic weight redundancy reduction loss is better than that of the baseline method. This is due to the effective extraction of cross-domain invariant causal features, which enables the model to be trained on a single source domain data with strong generalization performance. The M2 method adds ATWNNR to the M1 method, and the average accuracy on the two datasets is improved by 1.99% and 1.17% respectively compared with the M1 method. This shows that ATWNNR constrains the rank of causal features, making the extracted causal features more cross-domain invariant. The M3 method adds PGTL to the M1 method, and the average accuracy on the two datasets is improved by 3.64% and 2.39% respectively compared with the M1 method. This shows that PGTL effectively enhances the model's generalization ability to unknown working conditions by constraining the similarity between intra-class feature distribution and prototype center. The average accuracy of the proposed CDSSDGN method on the two datasets is 98.82% and 83.33% respectively, which is better than all ablation methods. The experimental results show that the method realizes the global optimization of feature representation through the multi-module collaborative optimization of the contrastive decoupling module, HSIC dynamic weight loss, ATWNNR rank constraint and PGTL prototype learning. Compared with a single module, the CDSSDGN method shows more stable performance and stronger generalization ability, verifying the effectiveness of the innovative design.

[0201] Although the present application discloses as above, the protection scope of the present application is not limited to this. The person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application, and these changes and modifications will fall within the protection scope of the present application.

[0202] The documents cited by the present application include:

[0203] [1]Qin Y, Yao Q, Wang Y, et al. Parameter sharing adversarial domain adaptation networks for fault transfer diagnosis of planetary gearboxes. Mechanical Systems and Signal Processing, 2021, 160: 107936.

[0204] [2]Xiao Y, Shao H, Lin J, et al. BCE-FL: A secure and privacy-preserving federated learning system for device fault diagnosis under Non-IID condition in IIoT. IEEE Internet of Things Journal, 2024, 11(8): 14241-14252.

[0205] [3]Xu G, Liu M, Jiang Z, et al. Online fault diagnosis method based on transfer convolutional neural networks. IEEE Transactions on Instrumentation and Measurement, 2020, 69(2): 509-520.

[0206] [4]Wang H, Xu J, Yan R, et al. A new intelligent bearing fault diagnosis method using SDP representation and SE-CNN. IEEE Transactions on Instrumentation and Measurement, 2019, 69(5): 2377-2389.

[0207] [5]Kang S, Tang X, Wang Y, et al. Cross-domain fault diagnosis method for rolling bearings based on contrastive universal domain adaptation. ISA transactions, 2024, 146: 195-207.

[0208] [6]Zhao C, Shen W. Mutual-assistance semisupervised domain generalization network for intelligent fault diagnosis under unseen working conditions. Mechanical Systems and Signal Processing, 2023, 189: 110074.

[0209] [7]Ragab M, Chen Z, Zhang W, et al. Conditional contrastive domain generalization for fault diagnosis. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 1-12.

[0210] [8] Ma H, Wei J, Zhang G, et al. Causality-inspired multi-sourcedomain generalization method for intelligent fault diagnosis under unknownoperating conditions. Reliability Engineering & System Safety, 2024, 252:110439.

[0211] [9] Wang R, Huang W, Lu Y, et al. A novel domain generalizationnetwork with multidomain specific auxiliary classifiers for machinery faultdiagnosis under unseen working conditions. Reliability Engineering & SystemSafety, 2023, 238: 109463.

[0212]

[10] Lei Li, Ke Gao, Juan Cao, et al. Progressive domain expansionnetwork for single domain generalization. In Proceedings of the IEEEConference on Computer Vision and Pattern Recognition (CVPR), 2021: 1-7.

[0213]

[11] Han T, Li Y F, Qian M. A hybrid generalization network forintelligent fault diagnosis of rotating machinery under unseen workingconditions. IEEE Transactions on Instrumentation and Measurement, 2021, 70:1-11.

[0214]

[12] Zhao C, Shen W. Adversarial mutual information-guided singledomain generalization network for intelligent fault diagnosis. IEEETransactions on Industrial Informatics, 2022, 19(3): 2909-2918.

[0215]

[13] Tang J, Ding X, Wei C, et al. HmmSeNet: A novel single domaingeneralization equipment fault diagnosis under unknown working speed usinghistogram matching mixup. IEEE Transactions on Industrial Informatics, 2024,20(5): 7162-7172.

[0216]

[14] Wang J, Ren H, Shen C, et al. Multi-scale style generative andadversarial contrastive networks for single domain generalization faultdiagnosis. Reliability Engineering & System Safety, 2024, 243: 109879.

[0217]

[15] Li G, Atoui M A, Li X. Dual adversarial and contrastive networkfor single-source domain generalization in fault diagnosis. AdvancedEngineering Informatics, 2025, 65: 103140.

[0218]

[16] Jia S, Li Y, Wang X, et al. Deep causal factorization network: A novel domain generalization method for cross-machine bearing fault diagnosis. Mechanical Systems and Signal Processing, 2023, 192: 110228.

[0219]

[17] Li J, Wang Y, Zi Y, et al. Causal consistency network: A collaborative multimachine generalization method for bearing fault diagnosis. IEEE Transactions on Industrial Informatics, 2022, 19(4): 5915-5924.

[0220]

[18] Jia L, Chow T W S, Yuan Y. Causal disentanglement domain generalization for time-series signal fault diagnosis. Neural Networks, 2024, 172: 106099.

[0221]

[19] Zhang G, Kong X, Ma H, et al. Dual disentanglement domain generalization method for rotating Machinery fault diagnosis. Mechanical Systems and Signal Processing, 2025, 228: 112460.

[0222]

[20] Guo C, Sun Y, Yu R, et al. Deep Causal Disentanglement Network with Domain Generalization for Cross-machine Bearing Fault Diagnosis[J]. IEEE Transactions on Instrumentation and Measurement, 2025.

[0223]

[21] An Y, Zhang K, Chai Y, et al. Domain adaptation network base on contrastive learning for bearings fault diagnosis under variable working conditions. Expert Systems with Applications, 2023, 212: 118802.

[0224]

[22] Liu Y, Wen W, Bai Y, et al. Self-supervised feature extraction via time–frequency contrast for intelligent fault diagnosis of rotating machinery. Measurement, 2023, 210: 112551.

[0225]

[23] Pan C, Shang Z, Tang L, et al. Open-set domain adaptive fault diagnosis based on supervised contrastive learning and a complementary weighted dual adversarial network. Mechanical Systems and Signal Processing, 2025, 222: 111780.

[0226]

[24] Li J, Wang Y, Zi Y, et al. Whitening-Net: A generalized network to diagnose the faults among different machines and conditions. IEEE transactions on neural networks and learning systems, 2021, 33(10): 5845-5858.

[0227]

[25] Miao Y, Zhao M, Lin J, et al. Application of an improved maximum correlated kurtosis deconvolution method for fault diagnosis of rolling element bearings. Mechanical Systems and Signal Processing, 2017, 92: 173-195.

[0228]

[26] Li C, Lei X, Huang Y, et al. Incrementally contrastive learning of homologous and interclass features for the fault diagnosis of rolling element bearings. IEEE Transactions on Industrial Informatics, 2023, 19(11): 11182-11191.

[0229]

[27] Zbontar J, Jing L, Misra I, et al. Barlow twins: Self-supervised learning via redundancy reduction. International conference on machine learning. PMLR, 2021: 12310-12320.

[0230]

[28] Zhang J, Lu J, Wang C, et al. Hyperspectral and multispectral image fusion via superpixel-based weighted nuclear norm minimization. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5521612.

[0231]

[29] Zhang L, Song L, Du B, et al. Nonlocal low-rank tensor completion for visual data. IEEE transactions on cybernetics, 2019, 51(2): 673-685.

[0232]

[30] Eren L, Ince T, Kiranyaz S. A generic intelligent bearing fault diagnosis system using compact adaptive 1D CNN classifier. Journal of Signal Processing Systems, 2019, 91(2): 179-189.

[0233]

[31] Pöppelbaum J, Chadha G S, Schwung A. Contrastive learning based self-supervised time-series analysis. Applied Soft Computing, 2022, 117:108397.

[0234]

[32] Smith W A, Randall R B. Rolling element bearing diagnostics using the Case Western Reserve University data: A benchmark study[J]. Mechanical Systems and Signal Processing, 2015, 64: 100-131.

[0235]

[33] Hong H S, Thuan N. HUST bearing: a practical dataset for ball bearing fault diagnosis[J]. Mendeley Data, 2023, 16: 138.

[0236]

[34] Kumar D, Mehran S, Shaikh M Z, et al. Triaxial bearing vibration dataset of induction motor under varying load conditions. Data in Brief, 2022, 42: 108315.

Claims

1. A rolling bearing fault diagnosis method based on comparative decoupling and single-source domain generalization, characterized in that, The method uses only one source domain data during the model training phase and includes the following steps: Step 1: Use the time-domain data of rolling bearings and perform two data augmentation processes on it to obtain two sets of augmented data; Step 2: The model is constructed based on a causal decoupling network and a contrastive learning BYOL framework. The causal decoupling network is used to extract causal and non-causal features of the data and constrains the rank of the causal features by adaptive threshold weighted nuclear norm regularization. The contrastive learning BYOL framework is used to bring the causal features extracted from the two sets of augmented data together to obtain the causal domain invariant features of each sample. Step 3: Use the prototype Gaussian triple loss function to constrain the causal domain invariant features, thereby improving the intra-class compactness and inter-class separability of the training samples; Step 4: Finally, input the target domain samples into the trained contrastive decoupling network model to achieve rolling bearing fault diagnosis.

2. The rolling bearing fault diagnosis method based on comparative decoupling and single-source domain generalization according to claim 1, characterized in that, Step one includes the following process: Using rolling bearing time-domain data, a time-domain sample set with a single sample length of L points is constructed by sliding sampling. Data augmentation processing such as adding noise, reversing signals, and window slicing is performed on each sample to obtain an augmented sample set.

3. The rolling bearing fault diagnosis method based on comparative decoupling and single-source domain generalization according to claim 1, characterized in that, Step two involves using a contrastive decoupling network to extract causal domain-invariant features, which includes the following process: Step 21: Separate causal features from non-causal features, specifically: First, the enhanced data samples are fed into a causal decoupling network, which introduces a structural causal model to describe the signal generation process: In the formula, Indicates to and fault labels They all possess causal characteristics of causal influence. Indicates impact Non-causal characteristics It can be considered as an unknown structure function. It is an unexplainable noise variable; Then, and Let represent the causal feature matrix and the non-causal feature matrix, respectively. The dynamic weight redundancy reduction loss (HSIC) can be expressed as: in, These are the weights of each item. This represents the HSIC operation, used to capture nonlinear redundancy; the three terms in the above formula respectively reduce internal redundancy of causal features, reduce internal redundancy of non-causal features, and eliminate nonlinear coupling between causal and non-causal features; weights The calculation method is expressed as follows: in, Used to calculate the approximate rank of the characteristic matrix. Used to calculate the singular values ​​of a matrix. , , , It is a temperature parameter; the HSIC calculation method is expressed as: in, Represents the L1 norm; and Both use radial basis function kernels. This is the core bandwidth parameter, with a value of 0.

5. , Represents the identity matrix. Represents a vector consisting entirely of 1s; After extracting features through a causal decoupling network and applying constraints from HSIC, causal and non-causal features can be obtained. Step 22: Constraining the rank of causal features, including the following process: For the causal feature matrix Singular value decomposition is performed to obtain each singular value. Calculate the normalized energy ratio: Set cumulative energy threshold (default The number of principal components is determined dynamically. : Design a piecewise weighting function for principal components and noise: Among them, principal component weights Smaller, reducing the impact on the previous one Singularity suppression strength, preserving key features; noise weighting Larger weights increase the penalty for tail singularities and suppress irrelevant components; the weights are increased. Combined with singular values, the adaptive threshold-weighted nuclear norm can be defined as: Using the reciprocal of the nuclear norm as a normalization term, minimizing the normalization term is equivalent to maximizing the nuclear norm; the adaptive regularized nuclear norm can be expressed as: In the formula, It is an adjustment factor. Indicates the number of samples; Steps two and three: Compare the decoupling models, including the following processes: The training process of the decoupled network is similar to that of the BYOL network, using two augmented samples from the original data. and The inputs are fed into two branches respectively, and then the upper and lower branches learn from each other to minimize the distance between the output features of the two branches. in, Indicates projector, Indicates the predictor, The dot product is a vector product used to measure the cosine similarity between two normalized vectors. Represents the L2 norm; To simultaneously optimize both branches of the network, the data from the upper and lower branches are swapped in the asymmetric prediction stage. The symmetric loss function is: 。 4. The rolling bearing fault diagnosis method based on comparative decoupling and single-source domain generalization according to claim 3, characterized in that, The prototype Gaussian triplet loss function constrained in step three, specifically, is as follows: in, It is a sample. It is a feature extraction network Output the features of the sample; The feature prototype is represented and updated using the EMA method, where... Indicates the iteration number at the epoch-th iteration. The prototype of each category, Indicates the first Sample size for each category ; It is the Gaussian kernel distance; It is right in front of you. It is a negative pair; This means that only the most difficult matching will result in a loss. This means that only the most difficult negative pairs will result in a loss; This is called the margin, used to ensure that the distance difference between positive and negative samples has a certain interval; the Gaussian kernel distance function is defined as: in, express and The Euclidean distance between them The standard deviation is defined by the formula: 。 5. The rolling bearing fault diagnosis method based on comparative decoupling and single-source domain generalization according to claim 4, characterized in that, Temperature parameters It takes values ​​between 0.3 and 0.

5.

6. A rolling bearing fault diagnosis system that compares and decouples single-source domain generalization, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 4 above, and executes the steps in the above-described comparative decoupling single-source domain generalization rolling bearing fault diagnosis method when running.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the comparative decoupling single-source-domain generalization rolling bearing fault diagnosis method as described in any one of claims 1 to 4.

Images