Class-incomplete multi-source-domain adaptive fault diagnosis method for gearbox
By generating high-fidelity pseudo-samples and performing feature alignment using the CWA-GAN and PCCAN frameworks, the class incompleteness problem in gearbox fault diagnosis is solved, achieving high accuracy and robust multi-source domain adaptive diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-27
AI Technical Summary
Existing multi-source domain adaptive methods suffer from class incompleteness in gearbox fault diagnosis, leading to negative transfer and information loss, making it difficult to effectively identify missing fault categories. Furthermore, the quality and distribution consistency of generated samples in generative adversarial networks are difficult to guarantee in class incomplete scenarios.
A high-fidelity pseudo-sample is generated using a Cyclic Consistency Wasserstein Generative Adversarial Network (CWA-GAN), and feature alignment is performed using a Prototype-Aware Class Conditional Adversarial Network (PCCAN), including global domain alignment, class conditional alignment, and prototype-based alignment. An end-to-end fault diagnosis framework is constructed, and the generation and alignment processes are optimized using a total loss function.
It effectively solves the problem of incomplete classes, generates high-quality pseudo samples, achieves multi-granularity feature alignment, and improves the accuracy and robustness of gearbox fault diagnosis, with an average accuracy of 98.81%.
Smart Images
Figure CN121743978A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, and more specifically, to a data-driven gearbox fault diagnosis method using deep learning, particularly a class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes. Background Technology
[0002] In recent years, with the development of sensor technology and artificial intelligence, deep learning-based fault diagnosis (DLFD) methods have been widely used in academia and industry due to their powerful automatic feature extraction capabilities. However, traditional deep learning methods are usually based on an idealized assumption: that the data used to train the model (source domain) and the data to be diagnosed in the actual application (target domain) follow the same probability distribution. In real industrial environments, this assumption is often difficult to hold. Because the working conditions of the gearbox, such as the operating load, speed, and ambient temperature, are constantly changing, the distribution of the collected vibration signals differs significantly. This phenomenon is called "domain shift." Domain shift causes the performance of a model trained under source domain conditions to deteriorate sharply under target domain conditions.
[0003] To address the domain offset problem, domain adaptation (DA) techniques emerged. Its core idea is to align the data distributions of the source and target domains during model training using specific algorithms (such as minimizing statistical distance or adversarial training), thereby improving the model's generalization ability in the target domain. However, early domain adaptation methods were mostly limited to single-source domain scenarios. In practical applications, building a robust diagnostic model often requires collecting multiple datasets from different times, devices, and operating conditions, which spurred the development of multi-source domain adaptation (MDA) techniques.
[0004] While multi-source domain adaptation techniques have made some progress, most existing methods implicitly rely on a stronger assumption: that all source and target domains share the exact same label space (i.e., the set of fault categories). However, in actual industrial data acquisition, due to limitations in equipment operating conditions or the randomness of fault occurrence, a particular source domain dataset may naturally lack one or more fault categories. For example, data acquired under operating condition A might only contain gearbox tooth fractures, tooth root cracks, and tooth wear, but not missing tooth faults. This "class incompleteness" problem poses a fatal challenge to existing multi-source domain adaptation methods. 1) Negative transfer problem: When a source domain with incomplete classes is forcibly aligned with a target domain with complete classes, the model will incorrectly match the feature of the missing class in the target domain with the feature of a class that exists in the source domain, resulting in severe class confusion and performance degradation, i.e., "negative transfer".
[0005] 2) Information gap problem: Since the model has never encountered the fault modes of the missing categories during the training phase, it is unable to learn the ability to identify these faults, resulting in it "ignoring" these fault categories in the target domain.
[0006] To address the issue of missing data, some studies have attempted to augment the data using techniques such as Generative Adversarial Networks (GANs). However, existing generative methods still face two main challenges when applied to multi-source domain scenarios with incomplete classes: First, how to ensure that the generated pseudo-samples retain the core semantic information of the fault while maintaining consistency with the distribution style of the target generation domain, especially under the complex mapping relationships of multi-source domains; second, after introducing generated samples, how to design an effective alignment mechanism that can simultaneously handle real and generated samples to achieve refined, category-aware feature alignment, thereby ensuring that the final features possess both domain invariance and high discriminative power.
[0007] Therefore, there is an urgent need for a new method that can systematically solve the problem of gearbox fault diagnosis in scenarios with incomplete multi-source domains. Summary of the Invention
[0008] To address the aforementioned problems and technical requirements, the inventors have proposed an adaptive fault diagnosis method for gearboxes with incomplete multi-source domains. The technical solution of this invention is as follows: An adaptive fault diagnosis method for gearboxes with incomplete multi-source domains specifically includes the following steps: Step 1: Under N different stable operating conditions of the gearbox, collect one-dimensional time-series vibration signals with multiple fault labels to form N source domain datasets, where at least one source domain dataset is class incomplete and N≥2; Under the actual operating conditions of the gearbox to be tested, collect unlabeled one-dimensional time-series vibration signals to form a target domain dataset.
[0009] Step 2: Perform data preprocessing and Fast Fourier Transform (FFT) on the N source domain datasets and target domain datasets to obtain frequency domain sample data for each domain; the data preprocessing includes: using a sliding window method to extract one-dimensional time series data samples from the vibration signal, and performing max-min normalization on the data samples to eliminate the influence of signal amplitude fluctuations.
[0010] Step 3: Train a Cyclic Consistent Wasserstein Generative Adversarial Network (CWA-GAN) using two source domains with complementary missing classes (e.g., source domain A and source domain B). The CWA-GAN contains two symmetric branches, each with a generator and a discriminator, to learn the bidirectional mapping relationship of the data distribution between the two source domains. Then, use the generator of the trained CWA-GAN to generate pseudo-samples of the missing class that conform to the style of each incomplete source domain. If source domain A is missing class k while source domain B has class k, then use the generator from domain B to domain A to convert the samples of class k in domain B into pseudo-samples in the style of domain A, thereby obtaining an augmented source domain dataset with N complete classes.
[0011] Step 4: Construct a prototype-aware class-conditional adversarial network (PCCAN) to achieve feature alignment and diagnosis of the data after category completion. PCCAN is trained end-to-end using N "enhanced source domain" datasets and a target domain dataset. PCCAN includes a shared feature extraction network, a fault classifier, and a multi-granularity feature alignment component. The shared feature extraction network extracts deep features from samples in all enhanced source and target domain datasets. The fault classifier predicts the fault category based on the extracted deep features. The multi-granularity feature alignment component aligns the extracted deep features through global domain alignment, class-conditional alignment, and prototype-based alignment. Global domain alignment can reduce the overall feature distribution differences between all augmentation source and target domains; class-conditional alignment: for each fault category, the feature distributions of the augmentation source and target domains are aligned separately for that specific category; prototype-based alignment: by explicitly constraining similar features to cluster towards their category center, intra-class compactness is enhanced.
[0012] Step 5: Input the test samples from the target domain dataset into the trained PCCAN feature extraction network and fault classifier. The output prediction results are the fault category identification and diagnosis results of the test samples.
[0013] The further technical solution is that, in step three, the training process of CWA-GAN is achieved by jointly optimizing a total loss function. The total loss function includes: adversarial loss based on Wasserstein distance and gradient penalty, which is used to improve the quality of generated pseudo samples and the stability of the training process; cycle consistency loss, which is used to ensure that the core information of the fault category of the sample remains unchanged during the generation of pseudo samples across domains; and identity mapping loss, which is used to maintain the domain-specific attributes of the sample during cross-domain transformation and prevent the model from over-transforming.
[0014] The further technical solution is that the adversarial loss based on Wasserstein distance and gradient penalty is defined as the loss function of the discriminator:
[0015] in, For the loss of the discriminator, To estimate the Wasserstein distance, a discriminator is trained to distinguish between real and generated samples by maximizing this term. This is a gradient penalty term used to impose a 1-Lipschitz constraint on the discriminator. This represents the gradient penalty coefficient.
[0016] The further technical solution is that the cycle consistency loss includes forward cycle consistency loss and backward cycle consistency loss. For the mapping between source domain A and source domain B with complementary missing categories, the forward cycle consistency loss penalizes the difference between a reconstructed sample from domain A to domain B and back to domain A and the original sample, and the backward cycle consistency loss penalizes the difference between a reconstructed sample from domain B to domain A and back to domain B and the original sample.
[0017] The further technical solution is that the global domain alignment in step four is achieved through a global domain discriminator. The global domain discriminator is trained to distinguish whether the input features come from any one of the N enhancement source domains or from the target domain. Furthermore, the gradient of the global domain discriminator is backpropagated through a gradient reversal layer (GRL) to train the feature extraction network adversarially, thereby prompting it to learn domain-invariant global features.
[0018] A further technical solution is that the class condition alignment in step four is achieved by configuring an independent binary local domain discriminator for each fault category, wherein, for the category... c The local discriminator is trained to distinguish a fault as predicted by the fault classifier. c The feature is whether it originates from the source domain or the target domain, and adversarial training is performed through the gradient inversion layer (GRL) to achieve fine alignment of the decision boundary for each class.
[0019] The further technical solution is that the prototype-based alignment in step four includes: maintaining a dynamically updated prototype vector in the feature space for each fault category. , Represents the center of the feature cluster of this category; the prototype alignment loss is defined for each source domain sample feature. Its corresponding category prototype vector The mean squared Euclidean distance between them, by minimizing the prototype alignment loss, drives the feature extraction network to learn a mapping that makes the features of samples of the same class spatially cluster toward their class prototypes.
[0020] Its further technical solution is, By calculating the current training batch Class center It is updated using an Exponential Moving Average (EMA) strategy, with the following update formula:
[0021] in, It is a category c prototype vector, It is the class center of the current batch of samples. It is the momentum coefficient, and .
[0022] A further technical solution is that, in step four, the training process of PCCAN is achieved by jointly optimizing a total loss function, which includes: classification loss on N augmentation source domain datasets. Global alignment loss caused by global domain alignment Local alignment loss caused by class-conditional alignment and the prototype alignment loss resulting from prototype-based alignment. Among them, the loss function Cross-entropy loss is used; loss function We employ weighted cross-entropy loss, where the weighting coefficients are determined by the fault classifier based on the class to which the sample belongs. c The posterior probability is determined; the total loss function is... The weighted sum of the above loss functions is expressed as:
[0023] in , , is the weighting coefficient for global alignment loss, local alignment loss, and prototype alignment loss.
[0024] A further technical solution is that the weighting coefficients of the global alignment loss, local alignment loss, and prototype alignment loss in the total loss function of PCCAN are adaptively and dynamically adjusted using a smooth ramp function during training to ensure that the classification task is learned first in the early stage of training, and then the strength of domain alignment is gradually increased.
[0025] The beneficial technical effects of this invention are: 1) This application systematically solves the problem of missing source domain categories. By using a data-level category completion framework (ICSC, the core of which is the CWA-GAN method), it eliminates the structural misalignment caused by missing categories from the source and effectively avoids negative transfer.
[0026] 2) This application generates high-fidelity, semantically consistent pseudo-samples. The designed CWA-GAN method combines Wasserstein adversarial loss, gradient penalty, cycle consistency loss and identity mapping loss to ensure high quality of generated samples, stable training, and preservation of core fault diagnosis semantics, and performs category completion at the data level.
[0027] 3) This application implements multi-granularity, structured feature alignment. The multi-granularity feature alignment and diagnosis (PCCAN) module is designed to achieve macro-domain invariance through a global domain discriminator; to achieve fine alignment of decision boundaries through a class-conditional local discriminator; and to enhance the intra-class compactness and inter-class separability of the feature space by applying explicit geometric constraints through prototype learning.
[0028] 4) This application has excellent diagnostic performance and robustness: Extensive validation results on the BJTU benchmark dataset show that the framework achieves an excellent average accuracy (98.81%), which confirms the practical value and advancement of the method in complex cross-domain diagnostic tasks. Attached Figure Description
[0029] Figure 1 This is a two-stage framework diagram of the class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes provided in this application.
[0030] Figure 2 This is a training architecture diagram of the ICSC framework provided in this application.
[0031] Figure 3 This is a structural diagram of the prototype-aware conditional adversarial network PCCAN provided in this application. Detailed Implementation
[0032] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0033] One embodiment of this application provides a class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes. The overall framework of this method includes two core stages: a collaborative data completion stage (ICSC) and a multi-granularity feature alignment and diagnosis stage (PCCAN). Figure 1 As shown, the specific implementation of this method includes the following steps: Step 1: Under N different stable operating conditions, acquire one-dimensional time-series vibration signals with multiple types of fault labels using signal acquisition equipment to form N source domains. (N≥2); each source domain Includes category labels y Vibration signal samples Its category set is Meanwhile, a large number of unlabeled vibration signal samples were collected under actual operating conditions (possibly variable speed conditions). This constitutes a target domain dataset. Its category set is C T The application scenario of this invention satisfies the condition that at least one source domain has an incomplete category, i.e. However, the union of all categories in the source domain equals the set of categories in the target domain, i.e. .
[0034] To verify the effectiveness of this method, this embodiment uses a publicly available vibration signal dataset: the Beijing Jiaotong University (BJTU) gearbox dataset for experiments. This dataset covers different equipment types, failure modes, speeds, and loads, providing a foundation for comprehensively evaluating the robustness of the model. The BJTU dataset includes five health states under three frequency conditions (20Hz, 25Hz, 30Hz): normal, tooth fracture, missing tooth, tooth root crack, and tooth wear.
[0035] To simulate real-world scenarios of class-incomplete multi-source domain adaptation (MDA), this embodiment constructs a series of cross-domain transfer tasks. Each task sets up two source domains and one target domain. Data in each source domain comes from specific operating conditions and is set to be class-incomplete, meaning that some fault category data is artificially removed from each source domain. Each source domain contains 100 labeled samples for each existing fault category. The target domain contains data for all fault categories and is used to evaluate the model's final diagnostic performance. In this way, transfer tasks comprehensively cover various combinations of operating conditions and class missing scenarios, ensuring a fair and rigorous evaluation of the method of this invention. Experiments were conducted on the BJTU dataset, constructing 18 different cross-domain transfer tasks to ensure comprehensive coverage of all possible class missing and domain shift scenarios.
[0036] Step Two: The dataset undergoes unified preprocessing. First, a sliding window approach is used to extract fixed-length (e.g., 2048 data points) one-dimensional time-series data sample segments from the original one-dimensional time-series vibration signal to ensure that the samples contain sufficient fault feature information. Then, each extracted data sample... x Perform max-min normalization to eliminate the impact of signal amplitude fluctuations caused by different operating conditions or sensor gain on feature extraction. The normalization operation is as follows:
[0037] in: The original vibration signal data, The samples after normalization. This represents the minimum value of each sample segment. This represents the maximum value of each sample segment. Finally, a Fast Fourier Transform is performed on the preprocessed data samples to obtain frequency domain feature data, which serves as the input for subsequent deep learning networks.
[0038] Step 3: The goal of the Collaborative Data Completion (ICSC) phase is to address the issue of incomplete source domain classes. This is achieved through data-level processing to resolve the common problem of incomplete class structures in gearbox scenarios. The core of the ICSC framework is the use of a specialized generative model—CWA-GAN (Cycle Wasserstein Augmentation Generative Adversarial Network)—to generate high-fidelity pseudo-samples for missing fault categories. For example... Figure 1 As shown in the left half, for any two source domains, such as source domain A and source domain B, the ICSC framework is used to complement each other's missing categories.
[0039] Reference Figure 2 As shown, the CWA-GAN network contains two symmetric branches, one of which is generated by the generator. and discriminator One branch is used to learn the mapping from domain A to domain B. Another branch consists of a generator. and discriminator The system is composed of components used to learn the mapping from domain B to domain A, thereby achieving a bidirectional mapping between domain A and domain B. This process is completed in two stages: CWA-GAN model training and pseudo-sample generation.
[0040] The training process employs the following optimization scheme:
[0041]
[0042] In the formula, and These represent the optimal generator network parameters after training, from source domain A to source domain B, and from source domain B to source domain A, respectively. and Let represent the optimal discriminator network parameters for source domain A and source domain B after training, respectively. The meanings of the other terms are explained below: and Let A and B represent the Wasserstein distance estimates, respectively. and These represent the gradient penalty terms for the discriminators in source domain A and source domain B, respectively. This represents the adversarial loss term of the generator. and These represent the forward and backward cycle-consistency losses for domains A and B, respectively. and These represent the identity mapping losses for domain A and domain B, respectively. , and represents the weight hyperparameters of the cycle consistency loss, identity mapping loss, and gradient penalty term, respectively.
[0043] 1) Wasserstein Adversarial Loss and Gradient Penalty: To overcome the instability of traditional GAN training, Wasserstein distance is used as a metric, and a gradient penalty (GP) mechanism is introduced to impose a 1-Lipschitz constraint on the discriminator, ensuring training stability and high-quality generated samples. The GP term is defined as follows:
[0044]
[0045] In the formula, Indicates the interpolation distribution The above interpolated samples Find the expected value. For discriminator Reconstructing samples from domain A The output, Similarly.
[0046] The corresponding Wasserstein distance estimate is:
[0047]
[0048] In each training iteration, the discriminator is updated n times to maximize the Wasserstein distance while satisfying the Lipschitz continuity constraint. This is equivalent to minimizing the following discriminator loss function:
[0049]
[0050] in, It is the gradient penalty coefficient.
[0051] The adversarial training process is constructed as a game between the generator and the discriminator. (Discriminator) and The generator is trained by maximizing the Wasserstein distance, thereby improving its ability to distinguish between real and generated samples. Conversely, the generator... and This distance is minimized by generating fake samples that can deceive the discriminator. Therefore, the adversarial loss function of the generator is defined as:
[0052] In the formula, This represents the true data distribution in source domain A. Above the sample Find the expected value. For discriminator Feed the input to the generator Domain B sample The output of the converted pseudo-sample, Similarly.
[0053] 2) Cyclic Consistency Loss: To ensure that the core diagnostic information of the sample is preserved during the cross-domain generation process, a cyclic consistency constraint is introduced. This loss penalizes the reconstruction error of a sample after cross-domain transformation and then reverse transformation, and is defined as follows:
[0054]
[0055] In the formula, , Represents the original samples of domain A and domain B. This represents a reconstructed sample that travels from domain A to domain B and back to domain A. This represents a reconstructed sample that travels from domain B to domain A and back to domain B.
[0056] 3) Identity Mapping Loss: To preserve the domain-specific attributes of samples during cross-domain transformations, an identity mapping loss is introduced. Specifically, when a sample comes from domain A... Input a generator from domain B to A At that time, the output results should be consistent with the original sample. Highly similar. Similarly, for samples from domain B... generator Should be maintained The inherent characteristic of this is that when a sample from one domain is input into a generator whose target domain is that domain, its output should be consistent with the input, defined as:
[0057]
[0058] In this embodiment, by jointly optimizing the Wasserstein adversarial loss, cycle consistency loss, and identity mapping loss, CWA-GAN can learn high-quality bidirectional mappings. After training, the CWA-GAN network can be used to complete data for all incomplete source domains. For example, if domain A is missing class k while domain B has class k, then all samples of class k in domain B will be included. Input into the trained generator In the process, pseudo-samples of domain A style are obtained. This bidirectional transformation process will be performed for all missing categories as follows: The generated pseudo-samples are added to the dataset of domain A, resulting in a fully categorical augmented source domain A. This process is repeated pairwise across all source domains to ultimately obtain N fully categorical augmented source domain datasets, where each augmented domain contains the original samples and their corresponding source-generated pseudo-fault samples. This bidirectional generation mechanism achieves complementarity in fault types, ensuring that both datasets are adequately populated in subsequent alignment stages.
[0059] In one possible implementation, the CWA-GAN network is trained for 100 epochs with a learning rate of 0.0001. The network structures of the generator and discriminator in CWA-GAN are shown in Table 1.
[0060] Table 1 CWA-GAN Network Parameters
[0061] Step 4: To achieve feature alignment and diagnosis of the completed data, a prototype-aware class-conditional adversarial network (PCCAN) is constructed to learn a domain-invariant and class-separable feature representation. (Refer to...) Figure 3 As shown, the PCCAN network includes the following core modules: 1) Feature extractor: a shared deep neural network (e.g., CA-MCNN based on convolutional neural networks) for extracting deep features from samples of all augmentation source and target domains. 2) Fault classifier: including a fully connected layer and a Softmax classifier for predicting fault categories based on the extracted deep features. 3) Three alignment components: a global discriminator (for global alignment), a local discriminator (for class-conditional alignment), and a prototype learning module (for prototype-based alignment). Samples from all N augmentation source and target domain datasets are first extracted into high-dimensional feature vectors by the shared feature extractor. Subsequently, the three alignment components work together: 3-1) Global Domain Discriminator: A multi-class discriminator used to distinguish whether a feature vector comes from one of the N augmentation source domains or from the target domain. Adversarial training is performed using a gradient inversion layer (GRL), forcing the feature extractor to generate features that the discriminator cannot distinguish, thus achieving macroscopic alignment.
[0062] 3-2) Local Domain Discriminator: A separate binary discriminator is configured for each fault category to distinguish whether the features of a specific category originate from the source domain or the target domain. Adversarial training using GRL enables the model to finely align the feature distribution within each category.
[0063] 3-3) Prototype Learning Module: For each fault category, a dynamically updated prototype vector is maintained in the feature space, representing the feature center of that category. Through a prototype alignment loss, samples of the same class are tightly pulled around their corresponding class prototype, thereby greatly enhancing the intra-class compactness of features.
[0064] The PCCAN network, constructed above, is trained end-to-end by inputting N augmented source domain datasets and target domain datasets. The training process involves jointly optimizing the source domain classification loss and the total loss function comprised of the three alignment losses mentioned above. This allows the model to learn feature representations that both eliminate domain gaps and form clear class structures. After training, high-precision unsupervised fault diagnosis can be performed on new target domain samples. The total loss function includes the following components: 1) Classification loss: Standard cross-entropy loss calculated on the augmented source domain dataset to ensure the class separability of features.
[0065] 2) Global Alignment Loss: Adversarial training is performed using a global discriminator and a gradient reversal layer (GRL) to minimize the cross-entropy loss of global domain classification, prompting the feature extractor to learn domain-invariant macroscopic features. The global alignment loss is defined as:
[0066] In the formula, This represents the total number of all source domain samples in the current training batch. This represents the total number of target domain samples in the current training batch. This represents the standard cross-entropy loss function, with the two parameters in parentheses representing the global discriminator's response to the first... The domain affiliation prediction probability output for the i-th sample, and the i-th The true domain labels of each sample.
[0067] 3) Class-Conditional Alignment Loss: Through adversarial training with C local discriminators and GRL, the weighted local classification loss is minimized to achieve fine-grained alignment of the decision boundaries for each class. The class-conditional alignment loss is defined as:
[0068] in, The source domain samples are predicted by the fault classifier. i Category c The probability, Similarly; and These are the outputs of the local discriminator for enhancing the features of the source and target domains, respectively.
[0069] 4) Prototype alignment loss: The prototype learning module provides a loss for each category. c Maintain a dynamic prototype vector P c ∈ ^D represents the center of the feature cluster for that category. These prototype vectors are calculated using batch-level centers from labeled source domain samples. µ c The exponential moving average (EMA) is updated. µ c Defined as:
[0070] in Indicates category c Mid-source domain samples ( The set of indices of ) The source domain sample features are used. Subsequently, the prototype vector is updated using the EMA strategy to ensure that the prototype vector stably and smoothly represents the historical feature centers of each category. The update formula is as follows:
[0071] To ensure intra-category compactness, the prototype alignment loss function is defined as: [The function is defined as follows:] ... Its corresponding category prototype vector Mean square of the Euclidean distance between them:
[0072] The total loss function is then the weighted sum of the above terms:
[0073] Weighting coefficients of each alignment loss Instead of being a fixed value, it is dynamically adjusted during the training process to achieve a smooth transition from initial feature learning to stable region alignment. In one possible implementation, a smooth ramp function is used during training. Adaptive adjustments are made to ensure that the focus is on the classification task in the early stages of training, and then the strength of domain alignment is gradually increased.
[0074]
[0075] In the formula, For the first i The preset maximum weight value for the alignment loss. This is the control parameter for the weighted growth rate. These are normalized training process variables.
[0076] During training, an alternating optimization strategy is used to update the parameters of each module, and the source domain classification loss is minimized jointly. Global combat losses Category-based conditional loss and prototype alignment loss This method achieves parameter optimization by progressively optimizing the feature space from coarse-grained global alignment to fine-grained category-aware representations with geometric structure.
[0077] In this embodiment, the PCCAN network includes a shared feature extractor based on CA-MCNN, a classifier, a global domain discriminator, C local domain discriminators, and a prototype learning module. The training epochs of the PCCAN network are set to 50 epochs, and the learning rate is set to 0.001. The weighting coefficients for the domain alignment loss are set to... , , Weighted growth rate control parameters Set to 5. The architecture configuration of each core module of PCCAN is shown in Table 2.
[0078] Table 2 PCCAN Model Parameters
[0079] Note: R, B, L, C, and K represent the set of real numbers (representing the feature dimension), training batch size, input signal length, total number of fault categories, and number of source domains, respectively.
[0080] Step 5: Evaluate the generalization performance of the trained model on the target domain: Unlabeled test samples from all categories in the target domain are input into the feature extractor and fault classifier of the PCCAN network trained in step four. The output is the final fault diagnosis category of the sample, and the classification accuracy of the model in the target domain is calculated. Verification shows that this embodiment achieves an average classification accuracy of 98.81% on the BJTU dataset, confirming the excellent robustness and diagnostic performance of the method in complex cross-domain diagnostic tasks.
[0081] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.
Claims
1. A class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes, characterized in that, The method includes the following steps: Step 1: Under N different stable operating conditions of the gearbox, collect one-dimensional time-series vibration signals with multiple fault labels to form N source domain datasets, where at least one of the source domain datasets is class incomplete and N≥2; Under the actual operating condition of the gearbox to be tested, collect one-dimensional time-series vibration signals without labels to form a target domain dataset. Step 2: Perform data preprocessing and fast Fourier transform on the N source domain datasets and the target domain dataset to obtain frequency domain sample data for each domain; Step 3: Train a Cyclic Consistency Wasserstein Generative Adversarial Network (CWA-GAN) using two source domains with complementary missing classes. The CWA-GAN contains two symmetrical branches, each with a generator and a discriminator, to learn the bidirectional mapping relationship of the data distribution between the two source domains. Then, use the generator of the trained CWA-GAN to generate pseudo-samples for each incomplete source domain that have the missing class and conform to the style of that source domain, to obtain an augmented source domain dataset with N complete classes. Step 4: Construct a prototype-aware class conditional adversarial network PCCAN to achieve feature alignment and diagnosis of the data after class completion, and perform end-to-end training of the PCCAN using N augmented source domain datasets and the target domain dataset; The PCCAN mentioned above includes: A shared feature extraction network is used to extract deep features from samples of all augmented source and target domain datasets; A fault classifier is used to predict fault categories based on extracted deep features; A multi-granularity feature alignment component is used to align extracted deep features through global domain alignment, class-conditional alignment, and prototype-based alignment. Step 5: Input the test samples from the target domain dataset into the trained PCCAN feature extraction network and fault classifier. The output prediction results are the fault category identification and diagnosis results of the test samples.
2. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 1, characterized in that, The global domain alignment in step four is achieved through a global domain discriminator, which is trained to distinguish whether the input features originate from any of the N augmentation source domains or from the target domain. Furthermore, the gradient of the global domain discriminator is backpropagated through a gradient inversion layer (GRL) to adversarially train the feature extraction network, thereby enabling it to learn domain-invariant global features and reduce the overall feature distribution differences between all augmentation source domains and the target domain.
3. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 1, characterized in that, The class condition alignment in step four is achieved by configuring an independent binary local domain discriminator for each fault category, wherein, for each category... c The local discriminator is trained to distinguish a fault predicted as a category by the fault classifier. c The feature is whether it originates from the source domain or the target domain, and adversarial training is performed through the gradient inversion layer (GRL) to achieve fine alignment of the decision boundary for each class.
4. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 1, characterized in that, The prototype-based alignment in step four includes maintaining a dynamically updated prototype vector in the feature space for each fault category. The Represents the center of the feature cluster of this category; the prototype alignment loss is defined for each source domain sample feature. Its corresponding category prototype vector The mean squared Euclidean distance between the features is used to drive the feature extraction network to learn a mapping that makes the features of samples of the same class spatially cluster toward their class prototypes by minimizing the prototype alignment loss, thereby enhancing intra-class compactness.
5. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 4, characterized in that, The By calculating the current training batch Class center It is updated using the exponential moving average (EMA) strategy, and the update formula is as follows: in, It is a category c prototype vector, It is the class center of the current batch of samples. It is the momentum coefficient, and .
6. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to any one of claims 1-5, characterized in that, In step four, the training process of PCCAN is achieved by jointly optimizing a total loss function, which includes: Classification loss on N augmented source domain datasets ; Global alignment loss caused by the global domain alignment Its loss function uses cross-entropy loss; Local alignment loss caused by the aforementioned conditional alignment Its loss function uses weighted cross-entropy loss, where the weighting coefficients are determined by the fault classifier based on the category to which the sample belongs. c The posterior probability is determined; Prototype alignment loss resulting from the prototype-based alignment The loss function is the mean squared Euclidean distance between each source domain sample feature and its corresponding class prototype vector. Total loss function The weighted sum of the above loss functions is expressed as: in , , is the weighting coefficient for global alignment loss, local alignment loss, and prototype alignment loss.
7. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 6, characterized in that, The weighting coefficients of the global alignment loss, local alignment loss, and prototype alignment loss in the total loss function are adaptively and dynamically adjusted during training using a smooth ramp function to ensure that the classification task is prioritized in the early stages of training, and then the strength of domain alignment is gradually increased.
8. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 1, characterized in that, In step three, the training process of CWA-GAN is achieved by jointly optimizing a total loss function, which includes: Adversarial loss based on Wasserstein distance and gradient penalty is used to improve the quality of generated pseudo samples and the stability of the training process; Cyclic consistency loss is used to ensure that the core information of the fault category of the sample remains unchanged during the process of generating pseudo samples across domains. Identity mapping loss is used to preserve the domain-specific attributes of samples during cross-domain transformations, preventing the model from over-transforming.
9. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 8, characterized in that, The adversarial loss based on Wasserstein distance and gradient penalty is defined as the loss function of the discriminator: in, The loss of the discriminator, To estimate the Wasserstein distance, a discriminator is trained to distinguish between real and generated samples by maximizing this term. This is a gradient penalty term used to impose a 1-Lipschitz constraint on the discriminator. This represents the gradient penalty coefficient.
10. The class-incomplete multi-source domain adaptive fault diagnosis method for gearboxes according to claim 8, characterized in that, The cycle consistency loss includes forward cycle consistency loss and backward cycle consistency loss. For a mapping between source domain A and source domain B with complementary missing classes, the forward cycle consistency loss penalizes the difference between a reconstructed sample from domain A to domain B and back to domain A and the original sample. The backward cycle consistency loss penalizes the difference between a reconstructed sample from domain B to domain A and back to domain B and the original sample.