Bearing fault diagnosis method, device and medium based on hierarchical collaborative domain adaptation

Abstract

The application relates to a bearing fault diagnosis method, device and medium based on hierarchical collaborative domain adaptation, the method comprising the following steps: acquiring bearing vibration signal samples for training and converting the bearing vibration signal samples into two-dimensional image representations to construct a training data set and input a bearing fault diagnosis model, the model comprising a feature extraction module, a hierarchical collaborative domain adaptation module and a fault classification branch, the feature extraction module is used for extracting shared deep features; the fault classification branch is used for classifying faults according to the shared deep features corresponding to source domain samples; the hierarchical collaborative domain adaptation module is used for sequentially performing global distribution alignment, category structure collaborative alignment and local topology preservation on source domain and target domain features; the bearing vibration signal of a bearing to be diagnosed is converted into a two-dimensional image representation, and the bearing vibration signal is processed through the trained feature extraction module and fault classification branch to obtain a bearing fault diagnosis result. Compared with the prior art, the application improves the accuracy and robustness of bearing fault migration diagnosis.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to bearing fault diagnosis methods, equipment and media based on hierarchical collaborative domain adaptation. Background Technology

[0002] Bearings are critical transmission components in rotating machinery, widely used in power, metallurgy, aerospace, and rail transportation industries. Their operating status directly affects the reliability, stability, and service life of the entire system. In actual industrial environments, rotating machinery typically operates under dynamic conditions with varying speeds and loads. The amplitude distribution, impact characteristic frequencies, and noise coupling characteristics of bearing vibration signals shift significantly with these changes in operating conditions, leading to a clear distribution difference between source domain data acquired under laboratory conditions and target domain data from the engineering field. If diagnostic models trained under fixed operating conditions are still used, problems such as feature representation failure, classification boundary shift, and decreased diagnostic accuracy are likely to occur. Meanwhile, labeled fault samples under new operating conditions in engineering fields are often difficult to obtain, and manual labeling is costly. Therefore, achieving stable and accurate cross-condition fault diagnosis under unlabeled conditions in the target domain has become a critical problem that urgently needs to be solved in the field of intelligent bearing diagnosis.

[0003] Deep transfer learning, by transferring knowledge from the source domain to the target domain, alleviates the problems of sample scarcity and distribution shift to some extent, and has become an important research direction for bearing fault diagnosis across operating conditions. Existing methods mostly focus on global domain adversarial and overall distribution alignment. While these methods can reduce the overall difference between the source and target domains, they still have the following shortcomings: First, single-scale feature extraction structures struggle to simultaneously consider local weak impact features, mesoscale periodic modulation features, and global operating condition distribution features, making it easy for key fault information to be weakened during the transfer process. Second, relying solely on overall distribution alignment without constraints on class structure can easily lead to aliasing of different fault categories in the feature space, thereby reducing the effectiveness of the target domain classification boundary. Third, the local manifold structure of target domain features is often not sufficiently constrained, potentially leading to increased intra-class discrepancies and blurred inter-class boundaries during the transfer process. Fourth, traditional domain discriminators often employ simple fully connected structures, which have limited ability to model local domain differences and structural shifts in deep features under complex operating conditions, thus limiting the effectiveness of global adversarial alignment. Fifth, the weights of multi-objective joint loss are usually set by human experience and lack the ability to adaptively adjust for different transfer tasks, which can easily lead to an imbalance in the optimization objectives.

[0004] Gram angular difference field (GADF) can convert one-dimensional vibration signals into two-dimensional image representations while preserving the temporal correlation of time series, providing a unified input format for convolutional neural networks. MobileNetV3 achieves a good balance between parameter scale and recognition accuracy through depthwise separable convolution and inverse residual structures, and has certain advantages in edge deployment. However, existing transfer diagnostic methods based on MobileNetV3 have not fully integrated multi-scale fault feature modeling, hierarchical collaborative domain adaptation, and enhanced domain discriminator design, making it difficult to effectively address cross-domain distribution migration problems under complex operating conditions.

[0005] Therefore, it is necessary to propose an unsupervised cross-condition bearing fault migration diagnosis method that takes into account multi-scale feature extraction, hierarchical collaborative domain adaptation, enhanced domain discriminative modeling, and loss weight adaptive optimization, so as to improve the diagnostic accuracy, robustness, and engineering applicability of the model under complex working conditions. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a bearing fault diagnosis method, device and medium based on hierarchical collaborative domain adaptation that takes into account multi-scale feature extraction, hierarchical collaborative domain adaptation, enhanced domain discriminative modeling and loss weight adaptive optimization, so as to achieve stable identification of bearing faults under the condition of no label in the target domain and improve the cross-domain adaptation capability and engineering deployment feasibility of the model.

[0007] The objective of this invention can be achieved through the following technical solutions: A bearing fault diagnosis method based on hierarchical collaborative domain adaptation includes: Obtain bearing vibration signal samples for training, which include source domain samples and target domain samples. Convert each bearing vibration signal sample into a two-dimensional image representation to construct a training dataset. The training dataset is input into a pre-built bearing fault diagnosis model for training. The bearing fault diagnosis model includes a feature extraction module, a hierarchical collaborative domain adaptation module, and a fault classification branch. The feature extraction module is used to extract bearing fault features at different scales based on the input two-dimensional image representation, and to perform adaptive weighted fusion and extraction of shared features to obtain shared deep features, which are then transmitted to the hierarchical collaborative domain adaptation module and the fault classification branch, respectively. The fault classification branch is used to classify faults based on the shared deep features corresponding to the source domain samples. The hierarchical collaborative domain adaptation module is used to perform global distribution alignment, category structure collaborative alignment, and local topology preservation on the shared deep features of the target domain samples to the shared deep features of the source domain samples in sequence, so that the fault classification branch can perform fault classification on the aligned shared deep features of the target domain samples. The bearing vibration signal of the bearing to be diagnosed is acquired and converted into a two-dimensional image representation. The two-dimensional image representation is then input into a trained bearing fault diagnosis model and processed through a feature extraction module and a fault classification branch to obtain the bearing fault diagnosis result.

[0008] Furthermore, the process of converting the bearing vibration signal into a two-dimensional image representation includes: The bearing vibration signal is mapped to the [-1,1] interval using the minimum-maximum normalization method, and combined with time information, the normalized vibration signal time series is obtained; The normalized vibration signal time series is mapped to a polar coordinate system, with the signal amplitude corresponding to the polar angle information and the time position corresponding to the radius information; Based on the vibration signal time series in polar coordinates, a two-dimensional GADF matrix is ​​calculated according to the angle difference, and the pixel values ​​are normalized to obtain a two-dimensional image representation.

[0009] Furthermore, the feature extraction module includes three parallel convolutional branches, a channel attention module, and a shared feature extraction backbone; The three parallel convolutional branches use convolutional layers with different dilation rates to construct differentiated receptive fields, so as to simultaneously extract local fine-grained impact features, mesoscale periodic modulation features and global operating condition distribution features of bearing faults from the two-dimensional image representation. The channel attention module is used to adaptively weight and fuse multiple features output from three parallel convolutional branches; The shared feature extraction backbone is used to extract shared features based on the weighted fusion results output by the channel attention module, and output shared deep features.

[0010] Furthermore, the global distribution alignment process performed by the hierarchical collaborative domain adaptation module includes: The shared deep features corresponding to the source domain samples and the target domain samples are extracted by a shared feature extractor. A lightweight convolutional domain discriminator is used to distinguish whether the shared deep features come from the source domain samples or the target domain samples, so as to align the shared deep features of the source domain samples and the target domain samples. During backpropagation, the gradient inversion layer is used to negate the propagated gradient. The lightweight convolutional augmentation domain discriminator includes a convolutional mapping unit, a residual augmentation unit, a global average pooling unit, and a domain classification unit connected in sequence. The convolutional mapping unit includes a convolutional layer and a ReLU activation function layer connected in sequence, which are used to perform convolution processing on the input feature map to extract local domain difference information and then output it. The residual enhancement unit includes a residual connection layer, a convolutional layer and a BN layer connected in sequence. It is used to perform residual processing on the feature map output by the convolutional mapping unit to enhance the feature transfer capability and improve the representation capability of structural domain shift before outputting. The global average pooling unit includes a global average pooling layer and a 1D vector output layer connected in sequence, which is used to perform average pooling on the feature map output by the residual enhancement unit and output a 1D vector. The domain classification unit comprises a fully connected layer, a Softmax activation function layer, and a domain classification layer connected in sequence. It is used to classify the 1D vector output by the global average pooling unit and output the domain classification result of the sample.

[0011] Furthermore, the hierarchical collaborative domain adaptation module's process for category structure collaborative alignment includes: By introducing the conditional maximum mean difference constraint, the mean embeddings of samples of each category in the source and target domains are calculated in the regeneration kernel Hilbert space, and a multi-scale conditional maximum mean difference loss is constructed to minimize the distance between similar conditional distributions, thereby enhancing the cross-domain aggregation capability of similar fault samples. The calculation expression for the mean embedding of samples of each category in the source and target domains is as follows: In the formula, For the first The mean embedding of class samples in the source domain samples. For the first The mean embedding of class samples in the target domain samples. and For the source domain samples and the target domain samples in the th... Number of samples in the class Represents the kernel mapping function. and They represent the elements belonging to the first digit in the source and target domains, respectively. A sample set of classes For the first sample in the source domain i One characteristic, For the first sample in the target domain j One feature; The calculation expression for the maximum mean difference loss under the multi-scale condition is as follows: In the formula, This represents the maximum mean difference loss under multi-scale conditions.

[0012] Furthermore, the local topology preservation process performed by the hierarchical cooperative domain adaptation module includes: By introducing MAC-LSC constraints, an adaptive adjacency graph is constructed based on high-confidence target domain samples, thereby dynamically correcting the adjacency graph and gradually optimizing the local manifold representation of the target domain; The MAC-LSC constraint establishes an adjacency weight matrix based on an adaptive adjacency graph. Based on the adjacency weight matrix, it promotes the clustering of similar nearest neighbor samples through intra-class compactness constraints and widens the gap between dissimilar nearest neighbor samples through inter-class separation constraints, thereby maintaining the consistency of the feature manifold structure of the target domain and improving the stability of the classification boundary. The expression for the adjacency weight matrix is: In the formula, For nodes in an adaptive adjacency graph j and nodes k The weights between them For nodes j The eigenvector at time t, For nodes k The eigenvector at time t, For scale parameters, For nodes j The neighborhood group, For nodes j Category prediction values, For nodes k The predicted values ​​for the categories; The expression for the intra-class compactness constraint is: In the formula, For intra-class compactness constraints; The expression for the inter-class separability constraint is: In the formula, For inter-class separation constraints, For the set of nearest neighbor samples of different classes, The minimum interval threshold; The expression for the MAC-LSC constraint is: In the formula, For MAC-LSC constraints, This is the balance coefficient.

[0013] Furthermore, the loss function in the training process of the bearing fault diagnosis model is a joint loss function, which is a weighted sum of classification loss, global adversarial loss, class structure co-alignment loss and local topology preservation loss. The classification loss is the cross-entropy classification loss of the fault classification branch for the shared deep features corresponding to the source domain samples. The global domain adversarial loss is the adversarial loss of the hierarchical collaborative domain adaptation module for global distribution alignment. The category structure collaborative alignment loss is the alignment loss of the hierarchical collaborative domain adaptation module for category structure collaborative alignment. The local topology preservation loss is the topology preservation loss of the hierarchical collaborative domain adaptation module for local topology preservation.

[0014] Furthermore, during the training process of the bearing fault diagnosis model, the joint loss function is used as the optimization objective. The backpropagation and iterative update of the bearing fault diagnosis model network parameters are completed through the Adam optimizer. Simultaneously, the weight coefficients of the joint loss function are optimized through Bayesian optimization iteratively until the bearing fault diagnosis model converges.

[0015] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs stored in the memory, wherein when the one or more programs are executed by the one or more processors, they implement a bearing fault diagnosis method based on hierarchical cooperative domain adaptation as described above.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements a bearing fault diagnosis method based on hierarchical cooperative domain adaptation as described above.

[0017] Compared with the prior art, the present invention has the following advantages: (1) Multi-scale feature enhancement to improve the integrity of fault representation: Local, mesoscale and global features of bearing faults are extracted synchronously through three parallel differential dilation rate convolution branches, which enhances the comprehensive representation capability of multi-level fault features under complex working conditions; and adaptive fusion is completed by combining channel attention mechanism, which strengthens the key feature response related to fault identification and suppresses redundant interference information.

[0018] (2) Hierarchical collaborative domain adaptation to improve the precision of cross-domain alignment: This invention constructs a hierarchical collaborative domain adaptation mechanism consisting of global distribution alignment, category structure collaborative alignment and local topology preservation. Through global distribution alignment, the overall distribution difference between the source domain and the target domain is suppressed; through category structure collaborative alignment, the cross-domain aggregation ability of similar fault samples is enhanced, and the category aliasing problem that may occur when relying solely on global alignment is alleviated; through local topology preservation, the consistency of the feature manifold structure of the target domain is maintained, and the stability of the classification boundary is improved. This allows cross-domain alignment to gradually extend from the overall distribution level to the category structure and local manifold level, thereby effectively reducing the risk of category aliasing and improving the stability of the target domain classification boundary.

[0019] (3) Adaptive Bayesian weights enhance task adaptability: This invention uses the Bayesian optimization method to globally optimize the weight coefficients in multi-objective loss, reducing the subjectivity caused by manual experience setting, and helping to improve the training stability and model adaptability under different cross-work condition transfer tasks.

[0020] (4) Lightweight convolution enhances the discriminator and improves the global adversarial effect: This invention improves the model’s ability to model local domain differences and structural domain shifts under complex conditions by replacing the traditional fully connected domain discriminator structure with convolutional mapping and small residual blocks, while avoiding the training burden caused by an overly heavy discriminator structure. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a bearing fault diagnosis method based on hierarchical collaborative domain adaptation provided in an embodiment of the present invention. Figure 2 This is a diagram of a bearing experimental apparatus provided in an embodiment of the present invention at the University of Paderborn, Germany. Figure 3 This is a schematic diagram of a GADF feature map construction process provided in an embodiment of the present invention; Figure 4 This is an overall architecture diagram of a bearing fault diagnosis model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the training accuracy curve when the target domain accounts for 20% in an embodiment of the present invention; Figure 6 This is a schematic diagram of the diagnostic accuracy curve under different target domain sample ratios to source domain sample ratios provided in an embodiment of the present invention. Figure 7 This is a schematic diagram comparing the accuracy of different migration task methods provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of a confusion matrix under an artificial damage scenario provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of a confusion matrix under a real damage scenario provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of t-SNE visualization results under an artificial injury scenario provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of t-SNE visualization results under a real damage scenario provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of a lightweight convolutional enhancement domain discriminator provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] Example 1 like Figure 1 As shown, this embodiment provides a bearing fault diagnosis method based on hierarchical collaborative domain adaptation, including the following steps: S1: Obtain bearing vibration signal samples for training. These samples include source domain samples and target domain samples. Convert each bearing vibration signal sample into a two-dimensional image representation to construct a training dataset. S2: Input the training dataset into the pre-built bearing fault diagnosis model for training. The bearing fault diagnosis model includes a feature extraction module, a hierarchical collaborative domain adaptation module, and a fault classification branch. The feature extraction module is used to extract bearing fault features at different scales based on the input two-dimensional image representation, and to perform adaptive weighted fusion and extraction of shared features to obtain shared deep features, which are then transmitted to the hierarchical collaborative domain adaptation module and the fault classification branch, respectively. The fault classification branch is used to classify faults based on the shared deep features corresponding to the source domain samples. The hierarchical collaborative domain adaptation module is used to perform global distribution alignment, category structure collaborative alignment, and local topology preservation on the shared deep features of the target domain samples to the shared deep features of the source domain samples in sequence, so that the fault classification branch can classify faults on the aligned shared deep features of the target domain samples. S3: Obtain the bearing vibration signal of the bearing to be diagnosed, and convert the bearing vibration signal into a two-dimensional image representation; input the two-dimensional image representation into the trained bearing fault diagnosis model, and process it through the feature extraction module and fault classification branch to obtain the bearing fault diagnosis result.

[0026] Specifically, in step S1, the process of converting the bearing vibration signal into a two-dimensional image representation includes: The Gram angular difference field GADF transform is used to convert the one-dimensional bearing vibration signal into a two-dimensional image representation in order to construct a unified network input.

[0027] In the GADF transform, the original bearing vibration signal is first normalized and mapped to the interval [-1,1]; then the normalized time series is mapped to the polar coordinate system; and finally, a two-dimensional image is constructed by the angle difference between any two sampling points.

[0028] The operation process in this embodiment includes the following sub-steps: S101: The bearing vibration signal is mapped to the [-1,1] interval using the minimum-maximum normalization method, and combined with time information, the normalized vibration signal time series is obtained to reduce the impact of signal amplitude scale differences under different working conditions. S102: Map the normalized vibration signal time series to the polar coordinate system, with the signal amplitude corresponding to the polar angle information and the time position corresponding to the radius information, thus fully preserving the temporal correlation of the time series; S103: Based on the vibration signal time series in polar coordinates, a two-dimensional GADF matrix is ​​calculated according to the angle difference, and the one-dimensional time series is converted into a two-dimensional image representation. The image size is determined according to the vibration signal length and the subsequent network input requirements, and the aspect ratio is kept consistent to avoid feature distortion.

[0029] S104: Map the pixel values ​​of the multi-channel feature images to the [0,1] range to avoid the numerical differences between channels causing gradient imbalance during network training, and ensure that the numerical range of each channel feature is compatible with the input requirements of the subsequent network.

[0030] The construction of a two-dimensional GADF matrix can completely preserve the fault characteristics of the bearing's inner ring, outer ring, and rolling elements, adapting to the feature input requirements for cross-operating condition migration diagnosis.

[0031] In step S2, the feature extraction module includes three parallel convolutional branches, a channel attention module, and a shared feature extraction backbone. The three parallel convolutional branches, combined with the channel attention mechanism, complete the adaptive weighted fusion of fault features under different receptive fields. Lightweight MobileNetV3 is used as the shared feature extraction backbone to realize hierarchical abstraction of fault features and shared deep feature extraction.

[0032] Specifically, the MobileNetV3 shared feature extraction backbone uses a multi-layered Bottleneck module to achieve feature abstraction from shallow to deep layers, providing a unified and stable deep representation for subsequent fault classification and domain adaptation.

[0033] Optionally, the three parallel convolutional branches use convolutional layers with different dilation rates to construct differentiated receptive fields, so as to simultaneously extract local fine-grained impact features, mesoscale periodic modulation features and global operating condition distribution features of bearing faults from the two-dimensional image representation. The channel attention module is used to adaptively weight and fuse multiple features output from three parallel convolutional branches; The shared feature extraction backbone is used to extract shared features based on the weighted fusion results output by the channel attention module, and outputs shared deep features.

[0034] In this embodiment, the three parallel convolutional branches are constructed with 3×3 convolutions with dilation rates of 1, 2 and 3 to create differentiated receptive fields, and simultaneously extract local fine-grained impact features, mesoscale periodic modulation features and global operating condition distribution features of bearing faults. The feature maps output from the three branches undergo adaptive weighted fusion via the channel attention module, which enhances the key feature responses related to fault identification, suppresses redundant interference information, and outputs enhanced feature maps. The shared feature extraction backbone is based on MobileNetV3 architecture. It utilizes its inverse residual structure, depthwise separable convolution and SE channel attention mechanism to complete the hierarchical abstraction of features through 11 Bottleneck modules. Finally, it outputs 1024-dimensional shared deep features through 1×1 convolution dimensionality increase and global average pooling, which serve as the common input for fault classification and hierarchical collaborative domain adaptation modules.

[0035] In step S2, the construction process of the hierarchical collaborative domain adaptation module includes: A hierarchical collaborative domain adaptation mechanism is constructed, consisting of global distribution alignment, class structure collaborative alignment, and local topology preservation, to achieve unsupervised cross-operating domain adaptation. The hierarchical collaborative domain adaptation module does not directly pass target domain features to the fault classification branch. Instead, it back-optimizes the feature extractor through multi-dimensional loss constraints, ensuring that the shared deep features of the target domain samples are distributed and aligned with the features of similar samples in the source domain within the same feature space. This allows the fault classification branch, trained only on labeled data in the source domain, to accurately classify the aligned target domain features.

[0036] Global distribution alignment optimizes domain adversarial loss and enhances the domain discriminator with lightweight convolution to improve the modeling ability of local domain differences and structural shifts under complex conditions. Class structure co-alignment minimizes the conditional distribution distance of similar samples in the source and target domains in the regenerating kernel Hilbert space through conditional maximum mean difference loss. Local topology preservation constructs an adaptive adjacency graph based on high-confidence pseudo-label samples in the target domain through local topology preservation loss and jointly implements intra-class local compactness constraints and inter-class boundary separation constraints.

[0037] Specifically, it includes: 1. Global Distribution Alignment: An adversarial learning architecture is constructed, consisting of a shared feature extractor, a gradient inversion layer, and a lightweight convolutional domain discriminator. Unlike traditional fully connected domain discriminators, this invention employs a lightweight convolutional domain discriminator. Two targeted innovative improvements are made to address the inherent shortcomings of traditional fully connected discriminators, such as losing feature space structure information and only capturing global statistical differences: First, it abandons the feature flattening operation of fully connected layers and uses convolutional mapping to directly process two-dimensional feature maps. It utilizes the local receptive field characteristics of convolution to extract local domain difference information in deep features region by region. This can accurately capture the subtle differences in local texture and amplitude distribution in bearing fault GADF images under different working conditions, solving the problem that fully connected structures cannot perceive local domain differences. Secondly, a small residual block specifically optimized for domain discrimination tasks is introduced within the discriminator. This design combines a bottleneck-type convolutional structure with an identity mapping shortcut, which not only solves the gradient vanishing problem during deep discriminator training and improves information transmission efficiency, but also models nonlinear structural domain shifts in the feature space caused by rotation and stretching due to varying speeds and loads. This overcomes the limitation of traditional fully connected structures, which can only model linear domain differences. Convolutional mapping enhances the perception of local domain differences in deep features, and the introduction of a small residual block within the discriminator improves information transmission efficiency and the ability to model structural domain shifts.

[0038] During training, the domain discriminator aims to distinguish features between the source and target domains, while the feature extractor learns domain-invariant representations through adversarial training, thereby suppressing the overall distribution differences between the source and target domains. The lightweight convolutional domain-enhanced discriminator comprises a convolutional mapping unit, a residual enhancement unit, a global average pooling unit, and a domain classification unit connected in sequence. The convolutional mapping unit is used to extract local domain difference information from shared deep features. The residual enhancement unit adopts a small residual block structure to enhance feature transfer capability and improve the representation capability of structural domain shifts. The domain classification unit is used to output the discrimination result of the domain to which the sample belongs.

[0039] like Figure 12 As shown, the lightweight convolutional enhancement domain discriminator adopts a structure that combines convolutional mapping and residual enhancement, which improves the ability to characterize local domain differences and structural domain shifts while controlling the parameter scale.

[0040] 2. Category Structure Collaborative Alignment: Based on global distribution coarse alignment, a conditional maximum mean difference constraint is introduced. The core idea is to measure the conditional distribution difference of each fault category in the source domain and the target domain in the regeneration kernel Hilbert space, and improve the category consistency of cross-domain feature alignment by minimizing the distribution distance between samples of the same type.

[0041] Let the deep discriminative features be The number of fault categories is For the first The class, source domain conditional distribution, and target domain conditional distribution are denoted as follows: and The category information of the target domain is provided by the high-confidence pseudo-labels output by the classifier. Conditional distribution alignment is implemented using deep features primarily because high-level representations are closer to the final classification semantics, and the clustering trend of similar samples is relatively more stable, making them more suitable for cross-domain distribution comparisons at the category level.

[0042] For the The mean embeddings of class samples in the reproducing kernel Hilbert space can be represented as follows: in, Represents the kernel mapping function. and They represent the elements belonging to the first digit in the source and target domains, respectively. A sample set of classes and This represents the number of samples for the corresponding category. After kernel mapping, the differences between the conditional distributions of different categories can be more clearly characterized in the high-dimensional feature space.

[0043] Therefore, the maximum mean difference loss under multi-scale conditions can be defined as: Expanding on this further, it can be written as: in, The kernel function is represented by the formula. This formula shows that the essence of CMMD is to minimize the mean difference between similar samples in the source and target domains at the category level, thereby enhancing the consistency of cross-domain features. Compared with simply relying on global domain adversarial approaches, by introducing multi-scale conditional maximum mean difference, the model no longer only focuses on whether the overall distributions of the source and target domains are similar, but further focuses on the correspondence between different fault categories in the feature spaces at each scale. This design can improve the discriminativeness of target domain features to a certain extent and mitigate the impact of category mismatch on the final diagnostic results. Especially when there are no manual labels in the target domain, implementing category constraints through high-confidence pseudo-labels can enhance the recoverability of the target domain category structure without increasing the labeling cost.

[0044] Compared to alignment methods that only focus on overall mean differences, this method further emphasizes the aggregation relationship of similar samples across operating conditions, thereby mitigating the class aliasing problem caused by global alignment. The mean embeddings of samples of each class in the source and target domains are calculated separately in the regenerating kernel Hilbert space, and the distance between distributions of similar conditions is minimized to enhance the cross-domain aggregation ability of similar fault samples. Target domain class information is provided by high-confidence pseudo-labels output by the classifier; this process helps mitigate the class aliasing problem that may occur when relying solely on global alignment.

[0045] 3. Local Topology Preservation: To maintain the local structural stability of the target domain's feature space, MAC-LSC constraints are introduced to construct an adaptive adjacency graph based on high-confidence target domain samples. This method constructs an adaptive adjacency graph based on high-confidence pseudo-label samples of the target domain. During the migration process, it constrains similar nearest neighbor samples to maintain a locally compact distribution and suppresses the disordered overlap of dissimilar samples in the local space, thereby enhancing the manifold consistency of the target domain's feature representation. Compared with the aforementioned CMMD, MAC-LSC focuses more on maintaining sample-level and neighborhood-level structural relationships; therefore, the two are complementary in terms of constraints.

[0046] For each high-confidence target domain sample, search for its nearest neighbor samples that match its pseudo-label within the current batch to form a local neighborhood set, and establish an adjacency weight matrix accordingly. The reason for using deep features to construct neighborhood relationships is that high-level representations have stronger category semantic information and can provide a more stable neighborhood structure foundation under pseudo-label conditions.

[0047] Adjacency weight matrix in, is the scale parameter. Based on the above adjacency relationship, a local compactness constraint can be further defined to encourage similar samples to form a higher degree of clustering in the local space. Its physical meaning is: if two target domain samples are determined to be of the same type and are neighbors at the current stage, they should be kept as close as possible during the transfer learning process, and should not be pulled away by disorder during the domain alignment process. Relying solely on local compactness constraints may lead to samples of different classes being excessively close in the boundary region. To address this, this paper further introduces a local separation term, imposing a margin constraint on sample pairs with different pseudo-labels but relatively close local distances to maintain the necessary class discrimination boundary. This constraint helps prevent dissimilar samples from the same class from being compressed into the same region during intra-class shrinkage, thereby improving inter-class separability in the target domain feature space. in, Represents the set of least neighbor samples of different classes. This represents the minimum margin threshold. Its role is to ensure sufficient spacing between local outlier samples, thereby forming a clearer decision boundary in the target domain feature space. This is particularly important for bearing fault diagnosis across operating conditions, as some categories may exhibit similar local patterns under changing operating conditions, easily leading to category confusion. Therefore, the total loss of MAC-LSC can be expressed as: in, The Ω is the balancing coefficient. Since the construction of neighborhood relationships depends on the current target domain feature distribution and pseudo-label confidence, MAC-LSC has strong adaptability, dynamically correcting the adjacency structure and gradually optimizing the local manifold representation of the target domain with training iterations. In other words, this constraint is not based on fixed neighborhood relationships and is completed all at once, but rather continuously updated along with the target domain features, making it more suitable for unlabeled cross-domain transfer scenarios.

[0048] By promoting the clustering of nearest neighbor samples of the same class through intra-class compactness constraints and widening the gap between nearest neighbor samples of different classes through inter-class separation constraints, the consistency of the feature manifold structure of the target domain is maintained and the stability of the classification boundary is improved.

[0049] In the hierarchical collaborative domain adaptation mechanism, both the category structure collaborative alignment constraint and the local topology preservation constraint are implemented based on the high-confidence pseudo-labels output by the classifier. The pseudo-label confidence threshold is set to 0.9, and only target domain samples with confidence higher than this threshold are selected to participate in the alignment process, so as to reduce the impact of pseudo-label error accumulation on training stability.

[0050] In step S2, the training process of the hierarchical collaborative domain adaptation module adopts Bayesian optimization-based adaptive optimization of the joint loss weights. Specifically, a joint loss function is constructed, consisting of classification loss, global domain adversarial loss, class structure collaborative alignment loss, and local topology preservation loss. Bayesian optimization is then used to adaptively optimize the weights of each loss term, including: The expression for the joint loss function is: in The cross-entropy classification loss is the cross-entropy of labeled samples in the source domain. , , These are the weight coefficients for global adversarial loss, category structure co-alignment loss, and local topology preservation loss, respectively.

[0051] Bayesian optimization search: , , Using the diagnostic accuracy of the validation set in the target domain as the objective function, a Bayesian optimization framework is constructed. First, a Gaussian process surrogate model is trained using initial sampling points. Then, the optimal weight combination is iteratively selected by balancing exploration and utilization through a sampling function. A two-layer optimization mechanism combining outer Bayesian optimization and inner gradient descent training achieves adaptive configuration of loss weights, reducing the subjectivity of manual parameter tuning.

[0052] Step S3 completes the training and online diagnostic deployment of the transfer diagnostic network, specifically including: Dataset partitioning: Bearing vibration signals under different working conditions were collected, and a labeled dataset in the source domain and an unlabeled dataset in the target domain were constructed. The datasets were then divided into training set, validation set and test set according to the proportions. Model Iterative Training: Input the training set samples into the network, take the joint loss function as the optimization objective, complete the backpropagation and iterative update of the network parameters through the Adam optimizer, and simultaneously optimize the loss weight coefficients through Bayesian optimization until the model converges and saves the optimal model weights. Online fault diagnosis: Real-time acquisition of bearing vibration signals under target operating conditions, completion of preprocessing and GADF feature map construction consistent with the training phase, and input into the optimal model after training; the model only uses the feature extraction backbone and fault classification branch to complete lightweight forward inference and output bearing fault type and health status diagnosis results.

[0053] The following is an experimental example of the above scheme: 1) Preparation of experimental data like Figure 2 As shown, experimental data were obtained using the bearing test bench at the University of Paderborn, Germany. The data includes artificially created damage samples and real damage samples generated by accelerated life tests, which can well reflect the failure characteristics in actual engineering. As shown in Table 1, the experiment selected five different operating conditions with different speeds, load torques, and radial forces, covering both artificial and real damage domains.

[0054] Data parameters: The length of a single sample is 1024 points, and the sampling frequency is 64kHz.

[0055] Fault types include four categories: normal, inner ring fault, outer ring fault, and rolling element fault.

[0056] As shown in Table 2, a total of 12 cross-condition migration tasks were set up, including two types of scenarios: migration between artificial damage domains and migration from artificial damage domains to real damage domains. In each task, the source domain is labeled data and the target domain is unlabeled data. The training set, validation set and test set are divided in a ratio of 7:2:1.

[0057] Table 1 Experimental data of PU bearings Table 2. Experimental Setup for Cross-Operating-Case Domain Migration 2) Implementation of specific diagnostic steps Step S1: Vibration signal preprocessing and GADF feature map construction The original bearing vibration signals from the PU dataset were zero-mean and amplitude normalized, and the samples were segmented using a sliding window. For each sample, a GADF transform was performed to convert the 1024-bit one-dimensional vibration signal into a 256×256 two-dimensional feature map. The pixel values ​​were then normalized to the [0,1] interval to complete the network input standardization process. The feature construction process is as follows: Figure 3 As shown.

[0058] Step S2: Multi-scale feature enhancement and shared feature extraction The GADF feature map is input into the multi-scale feature enhancement module, where local, mesoscale, and global fault features are extracted via three parallel dilated convolution branches. After adaptive fusion by the channel attention module, the enhanced features are input into the MobileNetV3 shared feature extraction backbone. A multi-layer Bottleneck module performs hierarchical abstraction of the features, ultimately outputting 1024-dimensional shared deep features, which are simultaneously fed into the fault classification branch and the hierarchical collaborative domain adaptation module. The overall network architecture is as follows: Figure 4 As shown.

[0059] Step S3: Calculation of constraint loss During network training, samples from both the source and target domains participate in forward propagation. The source domain shares deep features as input to the fault classification branch to calculate the cross-entropy classification loss; the source and target domain deep features are input to the hierarchical collaborative domain adaptation module, which sequentially calculates the global domain adversarial loss, the class structure collaborative alignment loss, and the local topology preservation loss. The global domain adversarial loss is output by a lightweight convolutional enhancement discriminator. Based on the current weight configuration, these losses are weighted and combined to form a joint optimization objective.

[0060] Step S4: Bayesian weight optimization A two-layer optimization mechanism combining outer Bayesian optimization and inner gradient training is adopted. In the inner training layer, the network parameters are updated through the Adam optimizer under the condition of fixed loss weights. In the outer optimization layer, the validation set diagnostic accuracy is used as the objective function. The optimal weight combination is selected through Bayesian optimization iteration, and finally the model parameters with the best performance on the validation set are saved.

[0061] Step S5: Cross-condition fault diagnosis reasoning and result analysis The target domain test samples are input into the trained optimal model. Forward inference is performed using the feature extraction backbone and fault classification branch to output the bearing fault type. The cross-condition transfer diagnostic performance of the model is comprehensively evaluated by combining the diagnostic accuracy, confusion matrix, and t-SNE feature visualization results.

[0062] 3) Experimental Results and Analysis Experimental results on the PU dataset demonstrate that this method achieves high diagnostic accuracy in the artificial damage domain transfer task. Figure 5 and Figure 6 As shown, the average diagnostic accuracy reached 97.2%, and samples of all categories were able to be distinguished relatively accurately. In the more challenging task of transferring from artificial damage domains to real damage domains, the model's average diagnostic accuracy remained above 92.5%, with only a few misclassifications occurring between similar local patterns such as outer ring faults and rolling element faults. Compared to the comparison methods, the accuracy improved by approximately 4% to 7%. Furthermore, as the proportion of target domain samples to source domain samples increased, the model's accuracy generally showed an upward trend, indicating that the constructed hierarchical collaborative domain adaptation mechanism has good transfer stability.

[0063] Depend on Figure 7 The comparison results for different transfer tasks show that, regardless of whether it's transferring between artificial damage domains or transferring from an artificial damage domain to a real damage domain, the proposed method maintains a high accuracy rate in target domain diagnosis, outperforming comparable methods such as DANN, DJDAN, and UDGAT. This indicates that multi-scale feature enhancement, lightweight convolutional enhancement of the domain discriminator, and hierarchical collaborative domain adaptation mechanisms can effectively improve cross-domain adaptability in complex scenarios.

[0064] like Figure 8 and Figure 9 As shown, in the artificial damage domain transfer task, samples of each category are mainly concentrated near the main diagonal. The four states—normal, inner circle, outer circle, and rolling body—can all be accurately distinguished, indicating that the model forms relatively clear category boundaries after cross-domain alignment within the same category. For the artificial damage domain to real damage domain transfer task, although the overall accuracy decreases, the main misclassifications are still concentrated between categories with similar local patterns, indicating that real damage samples have more complex signal distributions and higher intra-class dispersion. In this scenario, the proposed method still maintains good overall recognition performance, demonstrating the strong robustness of the proposed structure. Figure 10 and Figure 11The t-SNE visualization results further reveal that, after feature transfer learning, similar samples from the source and target domains exhibit a clear cross-domain aggregation trend in the feature space, forming relatively clear distribution boundaries between different fault categories. In the artificial damage domain task, the separation between various clusters is even higher; in the real damage domain task, although some clusters overlap, the overall clustering structure remains relatively clear. This indicates that the proposed method can effectively transfer fault features across different operating conditions.

[0065] 4) Experimental Conclusions Experimental results based on the PU bearing dataset demonstrate that the proposed method, through multi-scale feature enhancement, hierarchical collaborative domain adaptation mechanism, lightweight convolutional domain discriminator, and Bayesian weight adaptive optimization, can effectively reduce feature distribution shifts caused by changes in operating conditions under unlabeled target domain conditions, achieving high-precision cross-operating condition diagnosis of bearing faults. This method not only achieves good results in cross-operating condition transfer tasks on standard laboratory datasets but also maintains stable recognition performance in complex scenarios involving the transfer of artificial damage to real damage, demonstrating significant potential for engineering applications.

[0066] This embodiment also provides an electronic device, including: one or more processors; a memory; and one or more programs stored in the memory, wherein when the one or more programs are executed by one or more processors, they implement the bearing fault diagnosis method based on hierarchical cooperative domain adaptation as described above.

[0067] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the bearing fault diagnosis method based on hierarchical cooperative domain adaptation as described above.

[0068] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0069] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0073] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A bearing fault diagnosis method based on hierarchical collaborative domain adaptation, characterized in that, include: Obtain bearing vibration signal samples for training, which include source domain samples and target domain samples. Convert each bearing vibration signal sample into a two-dimensional image representation to construct a training dataset. The training dataset is input into a pre-built bearing fault diagnosis model for training. The bearing fault diagnosis model includes a feature extraction module, a hierarchical collaborative domain adaptation module, and a fault classification branch. The feature extraction module is used to extract bearing fault features at different scales based on the input two-dimensional image representation, and to perform adaptive weighted fusion and extraction of shared features to obtain shared deep features, which are then transmitted to the hierarchical collaborative domain adaptation module and the fault classification branch, respectively. The fault classification branch is used to classify faults based on the shared deep features corresponding to the source domain samples. The hierarchical collaborative domain adaptation module is used to perform global distribution alignment, category structure collaborative alignment, and local topology preservation on the shared deep features of the target domain samples to the shared deep features of the source domain samples in sequence, so that the fault classification branch can perform fault classification on the aligned shared deep features of the target domain samples. The bearing vibration signal of the bearing to be diagnosed is acquired and converted into a two-dimensional image representation. The two-dimensional image representation is then input into a trained bearing fault diagnosis model and processed through a feature extraction module and a fault classification branch to obtain the bearing fault diagnosis result.

2. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 1, characterized in that, The process of converting bearing vibration signals into a two-dimensional image representation by the method includes: The bearing vibration signal is mapped to the [-1,1] interval using the minimum-maximum normalization method, and combined with time information, the normalized vibration signal time series is obtained; The normalized vibration signal time series is mapped to a polar coordinate system, with the signal amplitude corresponding to the polar angle information and the time position corresponding to the radius information; Based on the vibration signal time series in polar coordinates, a two-dimensional GADF matrix is ​​calculated according to the angle difference, and the pixel values ​​are normalized to obtain a two-dimensional image representation.

3. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 1, characterized in that, The feature extraction module includes three parallel convolutional branches, a channel attention module, and a shared feature extraction backbone; The three parallel convolutional branches use convolutional layers with different dilation rates to construct differentiated receptive fields, so as to simultaneously extract local fine-grained impact features, mesoscale periodic modulation features and global operating condition distribution features of bearing faults from the two-dimensional image representation. The channel attention module is used to adaptively weight and fuse multiple features output from three parallel convolutional branches; The shared feature extraction backbone is used to extract shared features based on the weighted fusion results output by the channel attention module, and output shared deep features.

4. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 1, characterized in that, The process of global distribution alignment performed by the hierarchical collaborative domain adaptation module includes: The shared deep features corresponding to the source domain samples and the target domain samples are extracted by a shared feature extractor. A lightweight convolutional domain discriminator is used to distinguish whether the shared deep features come from the source domain samples or the target domain samples, so as to align the shared deep features of the source domain samples and the target domain samples. During backpropagation, the gradient inversion layer is used to negate the propagated gradient. The lightweight convolutional augmentation domain discriminator includes a convolutional mapping unit, a residual augmentation unit, a global average pooling unit, and a domain classification unit connected in sequence. The convolutional mapping unit includes a convolutional layer and a ReLU activation function layer connected in sequence, which are used to perform convolution processing on the input feature map to extract local domain difference information and then output it. The residual enhancement unit includes a residual connection layer, a convolutional layer and a BN layer connected in sequence. It is used to perform residual processing on the feature map output by the convolutional mapping unit to enhance the feature transfer capability and improve the representation capability of structural domain shift before outputting. The global average pooling unit includes a global average pooling layer and a 1D vector output layer connected in sequence, which is used to perform average pooling on the feature map output by the residual enhancement unit and output a 1D vector. The domain classification unit comprises a fully connected layer, a Softmax activation function layer, and a domain classification layer connected in sequence. It is used to classify the 1D vector output by the global average pooling unit and output the domain classification result of the sample.

5. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 1, characterized in that, The hierarchical collaborative domain adaptation module performs category structure collaborative alignment as follows: By introducing the conditional maximum mean difference constraint, the mean embeddings of samples of each category in the source and target domains are calculated in the regeneration kernel Hilbert space, and a multi-scale conditional maximum mean difference loss is constructed to minimize the distance between similar conditional distributions, thereby enhancing the cross-domain aggregation capability of similar fault samples. The calculation expression for the mean embedding of samples of each category in the source and target domains is as follows: In the formula, For the first The mean embedding of class samples in the source domain samples. For the first The mean embedding of class samples in the target domain samples. and For the source domain samples and the target domain samples in the th... Number of samples in the class Represents the kernel mapping function. and They represent the elements belonging to the first digit in the source and target domains, respectively. A sample set of classes For the first sample in the source domain i One characteristic, For the first sample in the target domain j One feature; The calculation expression for the maximum mean difference loss under the multi-scale condition is as follows: In the formula, This represents the maximum mean difference loss under multi-scale conditions.

6. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 1, characterized in that, The local topology preservation process of the hierarchical cooperative domain adaptation module includes: By introducing MAC-LSC constraints, an adaptive adjacency graph is constructed based on high-confidence target domain samples, thereby dynamically correcting the adjacency graph and gradually optimizing the local manifold representation of the target domain; The MAC-LSC constraint establishes an adjacency weight matrix based on an adaptive adjacency graph. Based on the adjacency weight matrix, it promotes the clustering of similar nearest neighbor samples through intra-class compactness constraints and widens the gap between dissimilar nearest neighbor samples through inter-class separation constraints, thereby maintaining the consistency of the feature manifold structure of the target domain and improving the stability of the classification boundary. The expression for the adjacency weight matrix is: In the formula, For nodes in an adaptive adjacency graph j and nodes k The weights between them For nodes j The eigenvector at time t, For nodes k The eigenvector at time t, For scale parameters, For nodes j The neighborhood group, For nodes j Category prediction values, For nodes k The predicted values ​​for the categories; The expression for the intra-class compactness constraint is: In the formula, For intra-class compactness constraints; The expression for the inter-class separability constraint is: In the formula, For inter-class separation constraints, For the set of nearest neighbor samples of different classes, The minimum interval threshold; The expression for the MAC-LSC constraint is: In the formula, For MAC-LSC constraints, This is the balance coefficient.

7. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 1, characterized in that, The loss function used in the training process of the bearing fault diagnosis model is a joint loss function, which is a weighted sum of classification loss, global adversarial loss, class structure co-alignment loss and local topology preservation loss. The classification loss is the cross-entropy classification loss of the fault classification branch for the shared deep features corresponding to the source domain samples. The global domain adversarial loss is the adversarial loss of the hierarchical collaborative domain adaptation module for global distribution alignment. The category structure collaborative alignment loss is the alignment loss of the hierarchical collaborative domain adaptation module for category structure collaborative alignment. The local topology preservation loss is the topology preservation loss of the hierarchical collaborative domain adaptation module for local topology preservation.

8. The bearing fault diagnosis method based on hierarchical collaborative domain adaptation according to claim 7, characterized in that, During the training process of the bearing fault diagnosis model, the joint loss function is used as the optimization objective. The backpropagation and iterative update of the bearing fault diagnosis model network parameters are completed through the Adam optimizer. At the same time, the weight coefficients of the joint loss function are optimized through Bayesian optimization iteratively until the bearing fault diagnosis model converges.

9. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs stored in the memory, which, when executed by the one or more processors, implement a bearing fault diagnosis method based on hierarchical cooperative domain adaptation as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements a bearing fault diagnosis method based on hierarchical collaborative domain adaptation as described in any one of claims 1 to 9.

Images