A cross-domain diagnosis method based on ternary feature decoupling

Abstract

This invention proposes a ternary feature decoupling network that uses a sliding window to process bearing data from the source and target domains, obtaining time-domain samples to form a dataset. The network includes a feature extractor, a decoupling branch module, a decoder, and a domain discriminator: the feature extractor extracts domain-invariant features; the decoupling branch module decouples the features into three independent subspaces: domain-invariant (strongly correlated with health status), domain-private (weakly correlated with health status), and random noise interference; the domain discriminator aligns the domain-invariant features of the source and target domains through adversarial training; and the decoder reconstructs the domain-invariant features and aligns them with the input samples. During training, classification loss, domain adversarial loss, reconstruction loss, and MMD alignment with a normal distribution constraint are used to achieve effective separation of cross-domain features. This invention avoids the learning of fault modes by nonlinear coupling interference in a unified feature space, improving the accuracy and generalization performance of fault diagnosis, and providing a new method for cross-domain intelligent diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis and relates to a cross-domain diagnostic method based on ternary feature decoupling. Background Technology

[0002] Bearings, gears, and shafts are fundamental components of machinery and are widely used in modern industry. Timely and accurate fault diagnosis of these components is crucial for ensuring the reliability of mechanical equipment and preventing dangerous accidents and damage.

[0003] Deep learning, with its powerful ability to extract features from massive datasets, offers a promising solution to this problem and has been widely applied in various emerging industrial applications. However, the performance of deep learning methods largely depends on two prerequisites: 1) sufficient labeled data; and 2) a uniform data distribution, which is often impractical in real-world applications. Specifically, data labeling requires specialized knowledge of the equipment, making it both time-consuming and labor-intensive, often resulting in insufficient labeled data. Furthermore, variations in operating conditions, mechanical structures, environmental noise, and other factors can lead to significant differences in the statistical properties of cross-domain data. This phenomenon, known as domain shift, typically degrades model performance when transferring knowledge from one domain to another. Therefore, unlabeled and domain-shifted data pose challenges to diagnostics. Given these issues, unsupervised domain adaptation (UDA) methods have found widespread application in diagnostics because they can reduce domain shift by transferring prior knowledge from the source domain to an unlabeled target domain extracted from a different but related distribution.

[0004] Despite significant progress in cross-domain fault diagnosis, existing UDA methods still suffer from key shortcomings that limit diagnostic performance improvement. In cross-device fault diagnosis, the original vibration signal typically contains three types of coupled feature components: domain-independent but strongly correlated with the health state, other private features that are domain-related but weakly correlated with the health state, and unavoidable random noise interference. These components exhibit complex nonlinear coupling relationships. Random noise interference not only severely reduces the stability and convergence speed of the diagnostic model but also masks true fault characteristics, leading to model misjudgment and reduced diagnostic accuracy and generalization ability. Other private features can easily cause the model to misidentify equipment differences as fault features, severely impacting the generalization ability of the diagnostic model. This nonlinear coupling severely interferes with the model's learning of true fault modes, reducing diagnostic accuracy and generalization performance. However, traditional binary decoupling methods (which only classify features into domain-invariant and domain-private categories) cannot effectively distinguish between useful domain-private information and harmful noise interference, resulting in the loss of valuable domain-private information while suppressing noise, or the introduction of noise pollution while retaining domain-private information. Therefore, accurately decoupling these three types of feature components, while retaining effective fault features and useful domain-private information, and fully suppressing multiple interferences from random noise, has become a key challenge in improving cross-device fault diagnosis performance. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a cross-domain diagnostic method based on ternary feature decoupling. This invention decouples the original feature representation into three subspaces with independent modeling value: first, discriminative information directly related to the fault type and shared across devices; second, domain-specific factors such as differences in device structure and changes in operating status; and third, irrelevant but unavoidable noise interference components. Through this complete feature decomposition, the model can focus on the alignment learning of domain-invariant features during the transfer process, avoiding ineffective fitting of domain-private features and noise features, thus improving the accuracy of cross-device diagnostics.

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

[0007] A cross-domain diagnostic method based on ternary feature decoupling includes the following steps:

[0008] S1: Use a sliding window approach to process long sequence data and construct time-domain signal samples of the source and target domains;

[0009] S2: Construct an adaptive network that includes a feature extractor, a decoupled branch module, a decoder, and a domain discriminator;

[0010] S3: Use trained feature extractors and classifiers to identify the health status of unlabeled data in the target domain;

[0011] Furthermore, step S2 specifically includes:

[0012] The ternary feature decoupling network feature extractor module is specifically as follows:

[0013] The feature extractor module in the ternary feature decoupling network consists of multiple cascaded convolutional layers, normalization layers, and activation functions. The input is vibration modal signal data from the source and target domains, and deep feature representations are extracted through a parameter-sharing feature extractor.

[0014] The decoupling branch module in the ternary feature decoupling network is specifically as follows:

[0015] By using the high-dimensional features output by the feature extractor as input, and decoupling the high-dimensional features through channel separation, the domain-invariant features φ corresponding to different domains of the three types of features are obtained. inv (x), other private features φ spec (x) and random noise interference characteristics φ noise (x), the source domain domain-invariant features are passed through the classification loss L CLS Supervised learning is performed, while random noise alignment is achieved through L... Noise Aligned with a normal distribution.

[0016] The decoder module of the ternary feature decoupling network is specifically as follows:

[0017] The decoder modules of the ternary feature decoupling network are a source domain feature decoder and a target domain feature decoder, which take the channel-fused source domain features and target domain features as input, specifically:

[0018] ;

[0019] Where, x src and x trc These represent the reconstructed samples from the source and target domains, respectively. The operator ⊕ indicates feature concatenation along the channel dimension. The decoder employs a reconstruction architecture combining multi-level upsampling and transposed convolution, achieving high-quality reconstruction from compressed features to the original signal through inverse operations. The decoder structure is mirror-symmetric to the feature extractor, using corresponding multi-level upsampling operations to gradually recover the signal length and utilizing transposed convolution for feature mapping and dimensionality upsampling.

[0020] The domain discriminator module in the ternary feature decoupling network is specifically as follows:

[0021] Source and target domain alignment is achieved through a domain discriminator. The input is the source and target domain invariant features extracted by the feature extractor. The domain discriminator determines whether the domain invariant features come from the source or target domain, and aligns the distribution between the two domains from a global perspective.

[0022] Furthermore, step S3 specifically includes:

[0023] Source Domain D s ={(x s_i , y s_i )|i=1,2,…,n s} and target domain sample D t ={(x t_j |j=1,2,…,n t As input, after decoupling by the channel separation module, three types of features are obtained: domain-invariant features φ corresponding to different domains. inv (x), other private features φ spec (x) and random noise interference characteristics φ noise (x).

[0024] The ternary feature decoupling network employs cross-entropy loss to supervise the learning of true fault labels on source domain samples, significantly improving the classifier's diagnostic accuracy in the source domain. The formula is as follows:

[0025] ;

[0026] Among them, y i,c and g i,c Let represent the label value of the i-th sample and its probability of belonging to class c, respectively. Simultaneously, domain adversarial loss is used to minimize the distribution difference between the source domain invariant features and the target domain invariant features, which can be expressed by the following formula:

[0027] ;

[0028] Where, d i Let R represent the binary label (domain label) of the i-th sample, and R() represent gradient inversion. Then, a set of reference noise samples following a standard normal distribution is constructed, and the maximum mean discrepancy (MMD) loss function is used to align the source domain noise features and target domain noise features decoupled from the model using these reference samples. This noise alignment loss function can be specifically expressed as:

[0029] ;

[0030] Where, φ s_noise (x s ) and φ t_noise (x t Z represents the noise features decoupled from the source and target domains, respectively. N(0,1) For the reference noise sample following a standard normal distribution, MMD(·) is the maximum mean difference measure function. To avoid the model focusing excessively on the reconstruction of samples from a single neighborhood, the ternary feature decoupling network reduces the source domain reconstruction loss L. src and target domain reconstruction loss L trcBy applying the L2 norm constraint, the total reconstruction loss L is obtained. RC Its expression is as follows:

[0031] ;

[0032] ;

[0033] ;

[0034] Where ||·||² represents the L2 norm. In summary, the complete loss function L of the ternary feature decoupling network... TDNN This can be expressed as follows:

[0035] ;

[0036] By training the model to convergence, we can effectively extract domain-invariant features and successfully transfer cross-domain knowledge, enabling the model to achieve classification accuracy in the target domain that is similar to that in the source domain, thereby solving the problem of distribution differences in domain adaptation.

[0037] The beneficial effects of this invention are as follows: This invention proposes a cross-domain diagnostic method based on ternary feature decoupling for fault diagnosis. Addressing the nonlinear coupling between three types of features—domain-independent but highly correlated with health status, domain-related but weakly correlated private features, and unavoidable random noise—the method constructs a ternary feature decoupling unit. Through a convolutional channel separation strategy and a three-branch structure, it effectively achieves cross-domain feature decoupling, thereby improving the accuracy and adaptability of fault diagnosis. This method can effectively achieve cross-operating condition fault diagnosis, overcoming the challenge of unlabeled fault samples, and providing a feasible solution for fault diagnosis. Specific advantages are as follows:

[0038] 1. This invention proposes a ternary feature decoupling network, which decomposes input data into three levels of features: random interference, domain invariant, and domain correlated features, using encoding reconstruction and channel feature separation strategies. Through joint constraints of domain adversarial loss, reconstruction loss, and normal distribution alignment loss, it achieves effective extraction of cross-domain invariant features, thereby improving diagnostic accuracy.

[0039] 2. This invention introduces a normal distribution alignment strategy to map and constrain random interference noise features to a standard Gaussian distribution space. Utilizing the prior knowledge and widespread presence of Gaussian noise in industrial environments, it effectively separates random noise components from nonlinearly coupled fault features. This not only reduces the interference of noise on the diagnostic model but also preserves the integrity and identifiability of the fault features.

[0040] Figure caption

[0041] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0042] Figure 1 This is a schematic diagram of the workflow framework for a cross-domain diagnostic method based on ternary feature decoupling; Detailed Implementation

[0043] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0044] This invention utilizes the Western Reserve University bearing dataset, the Hanoi University of Science and Technology bearing dataset, and bearing test bench data I and II built in the laboratory for verification. Detailed operating parameters and acquisition conditions for the four bearing datasets are shown in Table 1. Based on this, 12 cross-machine fault diagnosis tasks are constructed, specifically including: A→B, B→A, A→C, C→A, A→D, D→A, B→C, C→B, B→D, D→B, C→D, and D→C, where "→" indicates a fault diagnosis task from the source domain to the target domain.

[0045] Table 1 describes the cross-device experimental dataset and its operating conditions.

[0046]

[0047] As attached Figure 1 As shown, a cross-domain diagnostic method based on ternary feature decoupling includes the following steps:

[0048] S1: Use a sliding window method to process long sequence data and construct bearing time-domain signal samples in the source and target domains;

[0049] S2: Construct an adaptive network that includes a feature extractor, a decoupled branch module, a decoder, and a domain discriminator;

[0050] S3: Use trained feature extractors and classifiers to identify the bearing health status of unlabeled data in the target domain;

[0051] Furthermore, step S2 specifically includes:

[0052] The ternary feature decoupling network feature extractor module is specifically as follows:

[0053] The feature extractor module in the ternary feature decoupling network consists of multiple cascaded convolutional layers, normalization layers, and activation functions. The input is bearing vibration modal signal data from the source and target domains, and deep feature representations are extracted through a parameter-sharing feature extractor.

[0054] The decoupling branch module in the ternary feature decoupling network is specifically as follows:

[0055] By using the high-dimensional features output by the feature extractor as input, and decoupling the high-dimensional features through channel separation, the domain-invariant features φ corresponding to different domains of the three types of features are obtained. inv (x), other private features φ spec (x) and random noise interference characteristics φ noise (x), the source domain domain-invariant features are passed through the classification loss L CLS Supervised learning is performed, while random noise alignment is achieved through L... Noise Aligned with a normal distribution.

[0056] The decoder module of the ternary feature decoupling network is specifically as follows:

[0057] The decoder modules of the ternary feature decoupling network are a source domain feature decoder and a target domain feature decoder, which take the channel-fused source domain features and target domain features as input, specifically:

[0058] ;

[0059] Where, x src and x trc These represent the reconstructed samples from the source and target domains, respectively. The operator ⊕ indicates feature concatenation along the channel dimension. The decoder employs a reconstruction architecture combining multi-level upsampling and transposed convolution, achieving high-quality reconstruction from compressed features to the original signal through inverse operations. The decoder structure is mirror-symmetric to the feature extractor, using corresponding multi-level upsampling operations to gradually recover the signal length and utilizing transposed convolution for feature mapping and dimensionality upsampling.

[0060] The domain discriminator module in the ternary feature decoupling network is specifically as follows:

[0061] Source and target domain alignment is achieved through a domain discriminator. The input is the source and target domain invariant features extracted by the feature extractor. The domain discriminator determines whether the domain invariant features come from the source or target domain, and aligns the distribution between the two domains from a global perspective.

[0062] Furthermore, step S3 specifically includes:

[0063] Source Domain D s ={(x s_i , y s_i )|i=1,2,…,n s} and target domain sample D t ={(x t_j |j=1,2,…,n t As input, after decoupling by the channel separation module, three types of features are obtained: domain-invariant features φ corresponding to different domains. inv (x), other private features φ spec (x) and random noise interference characteristics φnoise (x).

[0064] The ternary feature decoupling network employs cross-entropy loss to supervise the learning of true fault labels on source domain samples, significantly improving the classifier's diagnostic accuracy in the source domain. The formula is as follows:

[0065] ;

[0066] Among them, y i,c and g i,c Let represent the label value of the i-th sample and its probability of belonging to class c, respectively. Simultaneously, domain adversarial loss is used to minimize the distribution difference between the source domain invariant features and the target domain invariant features, which can be expressed by the following formula:

[0067] ;

[0068] Where, d i Let R represent the binary label (domain label) of the i-th sample, and R() represent gradient inversion. Then, a set of reference noise samples following a standard normal distribution is constructed, and the maximum mean discrepancy (MMD) loss function is used to align the source domain noise features and target domain noise features decoupled from the model using these reference samples. This noise alignment loss function can be specifically expressed as:

[0069] ;

[0070] Where, φ s_noise (x s ) and φ t_noise (x t Z represents the noise features decoupled from the source and target domains, respectively. N(0,1) For the reference noise sample following a standard normal distribution, MMD(·) is the maximum mean difference measure function. To avoid the model focusing excessively on the reconstruction of samples from a single neighborhood, the ternary feature decoupling network reduces the source domain reconstruction loss L. src and target domain reconstruction loss L trc By applying the L2 norm constraint, the total reconstruction loss L is obtained. RC Its expression is as follows:

[0071] ;

[0072] ;

[0073] ;

[0074] Where ||·||² represents the L2 norm. In summary, the complete loss function L of the ternary feature decoupling network... TDNNThis can be expressed as follows:

[0075] ;

[0076] Labeled source domain sample data and unlabeled target domain sample data are imported together into a cross-domain diagnostic method based on ternary feature decoupling. The bearing health status of the unlabeled target domain data is identified by using the trained feature extractor and classifier to obtain the final test results.

[0077] To thoroughly verify the effectiveness of the proposed ternary feature decoupling method, a series of effectiveness verification experiments were designed. The core of this method lies in decoupling the feature space into three independent subspaces: domain-invariant features, domain-private features, and noise features, and employing differentiated processing strategies for each subspace. For ease of description, we name the noise feature channel Channel 1, the domain-invariant feature channel Channel 2, and the domain-private feature channel Channel 3. To comprehensively evaluate the contribution of each component in the ternary feature decoupling architecture, this paper designed the following two sets of comparative experiments: Channel 1 is removed, and the ratio of Channel 2 to Channel 3 is 3:1 (Method A); Channel 3 is removed, and the ratio of Channel 2 to Channel 1 is 3:1 (Method B).

[0078] The specific results are shown in Table 2. The experimental results show that, compared with the other two structures, the method of the present invention has improved accuracy in different cross-device experiments. This result verifies the effectiveness of the cross-domain diagnostic method based on ternary feature decoupling proposed in this invention.

[0079] Table 2 Results of cross-device experiments (%)

[0080]

[0081] Table 2 shows the experimental results of module effectiveness. The results indicate that the proposed complete ternary feature decoupling method performs best in all comparative experiments. Specifically, the original method achieves an average accuracy of 95.68% on the dataset, while the accuracy of method A drops to 89.05%, and the accuracy of method B is 94.13%. This demonstrates that the three feature subspaces are indispensable and together constitute a complete feature representation system.

[0082] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A cross-domain diagnostic method based on ternary feature decoupling, characterized in that: The method includes the following steps: S1: Use a sliding window approach to process long sequence data and construct time-domain signal samples of the source and target domains; S2: Construct an adaptive network that includes a feature extractor, a decoupled branch module, a decoder, and a domain discriminator; S3: Use trained feature extractors and classifiers to identify the health status of unlabeled data in the target domain.

2. The cross-domain diagnostic method based on ternary feature decoupling as described in claim 1, characterized in that: The ternary feature decoupling network feature extractor module is specifically as follows: The feature extractor module in the ternary feature decoupling network consists of multiple cascaded convolutional layers, normalization layers, and activation functions. The input is vibration modal signal data from the source and target domains, and deep feature representations are extracted through a parameter-sharing feature extractor.

3. The cross-domain diagnostic method based on ternary feature decoupling as described in claim 1, characterized in that: The decoupling branch module in the ternary feature decoupling network is specifically as follows: By using the high-dimensional features output by the feature extractor as input, and decoupling the high-dimensional features through channel separation, the domain-invariant features φ corresponding to different domains of the three types of features are obtained. inv (x), other private features φ spec (x) and random noise interference characteristics φ noise (x), the source domain domain-invariant features are passed through the classification loss L CLS Supervised learning is performed, while random noise alignment is achieved through L... Noise Aligned with a normal distribution.

4. The cross-domain diagnostic method based on ternary feature decoupling as described in claim 1, characterized in that: The decoder module of the ternary feature decoupling network is specifically as follows: The decoder modules of the ternary feature decoupling network are a source domain feature decoder and a target domain feature decoder, which take the channel-fused source domain features and target domain features as input, specifically: ； Where, x src and x trc These represent the reconstructed samples from the source and target domains, respectively. The operator ⊕ represents feature concatenation along the channel dimension. The decoder adopts a reconstruction architecture that combines multi-level upsampling with transposed convolution. It achieves high-quality reconstruction from compressed features to the original signal through inverse operations. The decoder structure forms a mirror symmetry with the feature extractor. It uses corresponding multi-level upsampling operations to gradually recover the signal length and uses transposed convolution for feature mapping and dimensionality upsampling.

5. The cross-domain diagnostic method based on ternary feature decoupling as described in claim 1, characterized in that: The domain discriminator module in the ternary feature decoupling network is specifically as follows: Source and target domain alignment is achieved through a domain discriminator. The input is the domain-invariant features of the source and target domains extracted by the feature extractor. The domain discriminator determines whether the domain-invariant features come from the source or target domain, and aligns the distribution between the two domains from a global perspective.

6. The cross-domain diagnostic method based on ternary feature decoupling as described in claim 1, characterized in that: The training steps of the cross-domain diagnostic method based on ternary feature decoupling are as follows: Source Domain D s ={(x s_i , y s_i )|i=1,2,…,n s } and target domain sample D t ={(x t_j |j=1,2,…,n t As input, after decoupling by the channel separation module, three types of features are obtained: domain-invariant features φ corresponding to different domains. inv (x), other private features φ spec (x) and random noise interference characteristics φ noise (x); The ternary feature decoupling network employs cross-entropy loss to supervise the learning of the true fault labels of source domain samples, thereby significantly improving the diagnostic accuracy of the classifier in the source domain. The formula is as follows: ； Among them, y i,c and g i,c Let represent the label value of the i-th sample and its probability of belonging to class c, respectively. Domain adversarial loss is used to minimize the distribution difference between the source domain invariant features and the target domain invariant features, which can be expressed by the following formula: ； Where, d i Let R represent the binary label (domain label) of the i-th sample, and R() represent gradient reversal. Then, a set of reference noise samples following a standard normal distribution is constructed, and the maximum mean discrepancy (MMD) loss function is used to align the source domain noise features and target domain noise features decoupled from the model using these reference samples. This noise alignment loss function can be specifically expressed as: ； Where, φ s_noise (x s ) and φ t_noise (x t Z represents the noise features decoupled from the source and target domains, respectively. N(0,1) For the reference noise samples that follow a standard normal distribution, MMD(·) is the maximum mean difference measure function; to avoid the model focusing too much on the reconstruction of samples from a single domain, the ternary feature decoupling network reduces the source domain reconstruction loss L. src and target domain reconstruction loss L trc By applying the L2 norm constraint, the total reconstruction loss L is obtained. RC Its expression is as follows: ； ； ； Where ||·||² represents the L2 norm; in summary, the complete loss function L of the ternary feature decoupling network is... TDNN This can be expressed as follows: ； By training the model to convergence, we can effectively extract domain-invariant features and successfully transfer cross-domain knowledge, enabling the model to achieve classification accuracy in the target domain that is similar to that in the source domain, thereby solving the problem of distribution differences in domain adaptation.

Images