Mask-guided feature decoupling based multi-source domain generalization cross-device fault diagnosis method and system

Abstract

The application belongs to but is not limited to the technical field of mechanical equipment fault diagnosis, and discloses a multi-source domain generalization cross-equipment fault diagnosis method and system guided by mask decoupling of features. The method generates domain-invariant masks and domain-specific masks through a mask generator, respectively, and applies the two masks to shallow features to obtain domain-invariant features and domain-specific features. Statistical dependence between the two masks is reduced through mutual information minimization loss, information loss in the decoupling process is reduced through feature reconstruction loss, and the fault discrimination ability of the features is maintained through a triple loss and a source domain classification loss. The training process does not need to use target domain equipment data, and the classifier outputs the fault category of unknown target equipment signals according to the domain-invariant features, thereby improving the generalization ability of cross-equipment fault diagnosis.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of mechanical equipment fault diagnosis technology, and particularly relates to a multi-source domain generalized cross-device fault diagnosis method and system with mask-guided feature decoupling. Background Technology

[0002] In modern industrial systems, machinery and equipment, as core assets, are widely deployed in key sectors such as petrochemicals, wind power generation, and rail transportation. Their operational status directly impacts production safety and economic efficiency. With the continuous improvement of industrial automation and intelligence, equipment is evolving towards larger scale, integration, and higher loads, leading to increasingly complex operating environments. Key components such as rolling bearings and gears operate under harsh conditions of variable speed, variable load, and high noise, making them highly susceptible to wear, cracks, and tooth breakage. However, due to significant differences in structural design, processing and assembly, sensor configuration, and operating environments among different devices, fault diagnosis models trained from data collected from a specific device (source domain) often suffer from a sharp decline in diagnostic performance, or even complete failure, when directly applied to another device (target domain), due to severe domain shift in data feature distribution. This problem is particularly prominent when the target equipment is newly installed or lacks historical fault records, severely restricting the cross-device generalization capability of intelligent diagnostic technology.

[0003] To address the aforementioned technical issues, existing technologies have proposed a class of fault diagnosis methods based on domain generalization. The core idea of ​​this class is to learn a general feature representation using data from multiple source domains, enabling it to directly generalize to a target domain never seen during training, without requiring any target domain data for adaptation. Existing methods mainly include data augmentation strategies (such as adding noise, temporal pruning, etc.) and feature alignment strategies (such as minimizing distribution differences between multiple source domains). However, these methods suffer from significant technical drawbacks: they mostly treat the distribution differences between different devices as a whole and weaken them, failing to effectively decouple device-independent fault characteristics from device-specific interference features. In practical cross-device diagnostic tasks, even for the same fault type, the vibration patterns presented on different devices often contain domain-specific information such as structural resonance and installation deviations. Existing methods cannot explicitly separate these two types of information in the feature space, resulting in the learned feature representation containing both fault discrimination information and residual device-specific interference. When faced with unknown target devices that differ greatly from the source domain, such mixed feature representations can lead to severe misclassification and degradation of generalization performance.

[0004] Therefore, how to effectively decouple domain-invariant fault features from domain-specific interference features by using only multiple labeled source domain data during the training phase, without the participation of any target domain data, and thus achieve accurate fault diagnosis of completely unknown devices, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the technical problems of poor generalization ability and difficulty in decoupling essential fault features from device interference features in cross-device fault diagnosis when the target device is completely invisible during the training phase, this invention provides a multi-source domain generalization cross-device fault diagnosis method with mask-guided feature decoupling. By integrating an intelligent random channel masking mechanism, mutual information minimization, and adaptive triplet loss, this method reduces information mixing between domain-invariant fault features and domain-specific interference features without requiring any target domain data, thereby improving the generalization ability and reliability of the diagnostic model on unknown target devices.

[0006] This invention is implemented as follows: a multi-source domain generalized cross-device fault diagnosis method with mask-guided feature decoupling, comprising the following steps:

[0007] Step 1: Obtain mechanical vibration signals and corresponding fault category labels from at least two source domain devices; perform sample division and standardization on the mechanical vibration signals to obtain multiple labeled source domain sample sets.

[0008] Step 2: Input the source domain samples into the feature extractor to extract shallow features, and construct a triplet consisting of anchor samples, positive samples, and negative samples based on the fault category labels of the source domain samples.

[0009] Step 3: Input the shallow features into the mask generator to generate a domain-invariant mask and a domain-specific mask respectively. Then, perform channel-wise weighted summation of the domain-invariant mask and the domain-specific mask with the shallow features to obtain the domain-invariant features and the domain-specific features respectively.

[0010] Step 4: Determine the mutual information minimization loss based on the domain-invariant mask and the domain-specific mask; determine the feature reconstruction loss based on the domain-invariant features, domain-specific features, and shallow features; determine the triplet loss based on the triplet; and determine the source domain classification loss based on the domain-invariant features and fault category labels. Train the feature extractor, mask generator, and classifier using a joint objective function that includes the source domain classification loss, triplet loss, mutual information minimization loss, and feature reconstruction loss, wherein the training process does not use target domain device data.

[0011] Step 5: Input the mechanical vibration signal to be diagnosed from the target domain device into the trained feature extractor and mask generator to obtain the domain-invariant features of the mechanical vibration signal to be diagnosed, and the classifier outputs the fault category based on the domain-invariant features.

[0012] Furthermore, the functions of each module in the model are as follows:

[0013] Shallow feature extractors are used to extract shallow convolutional features rich in temporal details from input source domain samples, avoiding excessive abstraction of device-specific information by deep networks;

[0014] The mask generator generates domain-invariant and domain-specific masks respectively, and decouples shallow features into domain-invariant and domain-specific features after channel-wise weighting.

[0015] A classifier, connected only to domain-invariant features, is used to classify source domain samples based on the decoupled domain-invariant features and predict fault categories.

[0016] Furthermore, the shallow feature extractor is constructed using the early convolutional layers of a one-dimensional deep shrinking residual network (1D-DRSN) to extract shallow features rich in temporal details, thus avoiding over-encoding of device-specific information by deep networks.

[0017] In this feature space, an anchor-positive-negative sample triplet sampling strategy is designed, and an adaptive boundary triplet loss is introduced to force the clustering of similar fault samples and the separation of dissimilar samples, thereby enhancing the discriminative ability of the features.

[0018] Furthermore, the mask generator receives the shallow features output by the shallow feature extractor and generates a domain-invariant mask and a domain-specific mask respectively; the domain-invariant mask is weighted channel-by-channel with the shallow features to obtain the domain-invariant features, and the domain-specific mask is weighted channel-by-channel with the shallow features to obtain the domain-specific features.

[0019] Furthermore, the mask generator is implemented as follows:

[0020] The input shallow features are processed by convolution and batch normalization, then introduced into a random perturbation layer. A Softmax layer maps the intermediate features to a channel probability distribution, followed by convolution, batch normalization, pooling, and activation. Two sub-networks output domain-invariant and domain-specific masks, respectively. These masks are not pre-defined as strictly complementary across channels, and are weighted across channels with the shallow features to obtain domain-invariant and domain-specific features. .

[0021] Furthermore, the losses in the joint objective function described in step 4 are defined as follows:

[0022] The source domain classification loss uses cross-entropy loss to supervise the training of the shallow feature extractor and classifier, enabling them to accurately identify fault categories in the source domain.

[0023] The adaptive triplet loss is used to construct anchor-positive-negative sample triplets in the shallow feature space, forcing the clustering of similar fault samples and the separation of dissimilar samples, thereby enhancing the discriminative ability of the features. Its boundary parameters are dynamically adjusted according to the anchor category.

[0024] To minimize the mutual information loss, a mutual information neural estimation framework is used to estimate the mutual information between a domain-invariant mask and a domain-specific mask, thereby reducing the statistical dependence between the two masks and prompting them to form differentiated channel allocation methods.

[0025] The feature reconstruction loss uses the Frobenius norm to measure the difference between the reconstructed features obtained by combining domain-invariant features and domain-specific features and the original shallow features, so that the reconstructed features approximate the original shallow features.

[0026] Another objective of this invention is to provide a multi-source domain generalized cross-device fault diagnosis system with mask-guided feature decoupling, characterized in that it includes:

[0027] Multi-source domain data acquisition module, used to acquire tagged mechanical vibration signals from multiple devices with different source domains;

[0028] The data preprocessing and enhancement module is used to slice, standardize, and add Gaussian perturbation noise to the vibration signal to form a source domain enhanced sample set;

[0029] The shallow feature extraction and triplet enhancement module is used to construct a one-dimensional deep shrinkage residual network to extract shallow features and calculate the adaptive triplet loss.

[0030] The mask generation and feature decoupling module is used to generate domain-invariant masks and domain-specific masks, decoupling shallow features into domain-invariant features and domain-specific features.

[0031] The mutual information minimization module is used to estimate the mutual information between a domain-invariant mask and a domain-specific mask, and to generate a mutual information minimization loss to reduce the statistical dependence of the two masks.

[0032] The feature reconstruction module is used to calculate the reconstruction loss between the original shallow features and the reconstructed features.

[0033] The classification module, connected only to domain-invariant features, is used to predict fault categories;

[0034] The model training module is used to input multiple source domain sample sets into the model and perform end-to-end training on the model by jointly optimizing the source domain classification loss, adaptive triplet loss, mutual information minimization loss and feature reconstruction loss.

[0035] The fault diagnosis module receives the trained model, processes the target domain test data that has never been seen before, and outputs the final fault category.

[0036] Another object of the present invention is to provide a computer device including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the mask-guided feature decoupling multi-source domain generalization cross-device fault diagnosis method.

[0037] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the mask-guided feature decoupling multi-source domain generalized cross-device fault diagnosis method.

[0038] Another objective of this invention is to provide an information data processing terminal, which includes the aforementioned mask-guided feature decoupling multi-source domain generalized cross-device fault diagnosis system.

[0039] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0040] This invention achieves explicit decoupling between domain-invariant fault features and domain-specific interference features through the design of an intelligent random channel masking mechanism. This design solves the core technical challenge of existing domain generalization methods, which struggle to separate equipment-related variations from the essential semantics of faults, leading to feature confusion and severe degradation of generalization performance under extreme operating conditions. It fundamentally improves the model's feature discrimination capability and diagnostic accuracy on completely unknown equipment.

[0041] To address the problem that existing methods fail to effectively utilize discriminative structural information in multi-source domain data, this invention introduces an adaptive boundary triplet loss. By constructing feature geometric constraints that cluster similar features and separate dissimilar features, the discriminative power of domain-invariant features is enhanced. Simultaneously, by combining mutual information minimization loss and feature reconstruction loss, effective information from the original shallow features is preserved while reducing the statistical dependence between domain-invariant and domain-specific masks, thus promoting differentiated information division between the two types of features.

[0042] This invention employs a fully unsupervised domain generalization paradigm, requiring no data from the target domain (including unlabeled data) throughout the diagnostic process. It achieves high-precision, zero-sample diagnosis of completely unknown devices using only data from multiple labeled source domains. This feature overcomes the technical bottleneck of traditional intelligent diagnostic models, which must rely on target domain data for adaptation or labeling when deploying new equipment. It significantly reduces the deployment cost and application threshold of intelligent operation and maintenance systems, making it particularly suitable for newly installed equipment or industrial scenarios lacking historical data.

[0043] This invention promotes the separation of fault discrimination information from equipment-related information by synergistically optimizing four mechanisms: mask-guided feature decoupling, adaptive triplet discrimination, mutual information minimization, and feature reconstruction. This reduces information loss during feature decoupling, thereby improving the generalization ability and robustness of cross-device fault diagnosis. In addition to being applicable to mechanical equipment health monitoring, the feature decoupling framework can also be used for other signal recognition tasks with different domain distributions.

[0044] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0045] The technical solution of this invention achieves multi-source domain generalized fault diagnosis through mask-guided feature decoupling. It can be deployed to entirely new equipment without requiring any target domain data, significantly reducing the implementation threshold and data dependence of industrial intelligent operation and maintenance. Expected benefits include:

[0046] Significantly reduced deployment costs: Traditional methods require collecting a large amount of labeled or even unlabeled data on new equipment for adaptation. This solution enables training once and deployment on multiple devices, reducing the deployment cost of a single new device by more than 90%. It is especially suitable for scenarios where fault data is difficult to obtain, such as wind power, rail transit, and aero engines.

[0047] Improved operational efficiency: No downtime is required to wait for data collection and model adaptation, enabling plug-and-play zero-sample diagnosis and reducing the mean time to repair (MTTR) of equipment by more than 30%.

[0048] Derivative commercial value: It can build a cross-device universal fault diagnosis cloud platform, provide services to multiple manufacturing enterprises through API subscription, form standardized diagnostic products, and expect annual service revenue to exceed 10 million yuan.

[0049] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0050] Breaking through the limitations of existing industrial internet platforms (such as GE Predix, Siemens MindSphere, and Rootcloud) that only support the same type of equipment or require target domain data adaptation, it achieves multi-source domain generalization diagnosis for the first time without requiring target domain data.

[0051] It fills the gap in the national standard GB / T 40571—2021 "General Requirements for Predictive Maintenance of Intelligent Services" for fault diagnosis technology with zero target sample data, and provides a brand-new generalization capability evaluation index for intelligent manufacturing remote operation and maintenance systems.

[0052] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0053] Feature decoupling challenge: Traditional domain generalization methods (such as DANN, MMD, and CORAL) cannot separate inherent interference and fault characteristics of devices, resulting in a diagnostic accuracy of only about 20% in scenarios with significant device differences. This invention improves the average accuracy to over 50% and the relative improvement to over 30% through explicit decoupling guided by masking and minimization of mutual information.

[0054] Zero-sample cold start problem: The industry has long desired that newly installed equipment could have diagnostic capabilities without any historical data, but existing methods rely on a large amount of source domain data plus a small amount of target domain data (or even unlabeled data) for adaptation. This invention achieves an average accuracy of 43.9% to 52.4% across multiple cross-device tasks under conditions of no target domain data, realizing a technological breakthrough in zero-sample cold start.

[0055] (4) The technical solution of the present invention overcomes technical bias:

[0056] Breaking the industry misconception that domain generalization performance is far inferior to domain adaptation: Traditional views hold that not using target domain data inevitably leads to significantly lower performance compared to domain adaptation methods. Experiments on four publicly available datasets demonstrate that MGFD's average accuracy (52.4%) is close to that of some domain adaptation methods (such as unsupervised CDAN, approximately 60%). However, the latter requires unlabeled target domain data, while this invention requires no target domain data at all, significantly improving engineering usability.

[0057] Breaking the inherent bias that deeper features are more conducive to generalization: Most methods assume that the deeper the network, the more abstract and generalized the extracted features. This invention takes the opposite approach, using shallow features for decoupling, avoiding the over-encoding of device-specific information by deep networks. Experiments show that shallow feature + mask decoupling is superior to direct classification by deep features, providing a new approach for domain-generalized feature extraction. Attached Figure Description

[0058] Figure 1 This is a flowchart of a multi-source domain generalized cross-device fault diagnosis method for mask-guided feature decoupling provided in an embodiment of the present invention;

[0059] Figure 2 A detailed flowchart of the multi-source domain generalized cross-device fault diagnosis method for mask-guided feature decoupling provided in this embodiment of the invention;

[0060] Figure 3 This is a structural diagram of the fault diagnosis system provided in an embodiment of the present invention;

[0061] Figure 4 This is a comparison chart of diagnostic results of different models under the dual-source domain generalization task provided in this embodiment of the invention;

[0062] Figure 5This is a comparison chart of diagnostic results of different models under the three-source domain generalization task (with the Shandong University of Science and Technology bearing dataset as the target domain) provided in the embodiments of the present invention;

[0063] Figure 6 This is a comparison chart of diagnostic results of different models under the three-source domain generalization task (with the Case Western Reserve University bearing dataset as the target domain) provided in the embodiments of the present invention;

[0064] Figure 7 This is a comparison chart of diagnostic results of different models under the three-source domain generalization task (with the Jiangnan University bearing dataset as the target domain) provided in the embodiments of the present invention;

[0065] Figure 8 This is a comparison chart of diagnostic results of different models under the three-source domain generalization task (with the Southeast University bearing dataset as the target domain) provided in the embodiments of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0067] like Figure 1 As shown in the figure, the multi-source domain generalized cross-device fault diagnosis method provided by the embodiment of the present invention for mask-guided feature decoupling includes the following steps:

[0068] S101, acquire mechanical vibration signals and corresponding fault category labels of at least two source domain devices, perform sample division and standardization processing on the mechanical vibration signals, and obtain multiple labeled source domain sample sets;

[0069] S102, the source domain samples are input into the feature extractor to extract shallow features, and a triplet consisting of anchor samples, positive samples and negative samples is constructed based on the fault category labels of the source domain samples.

[0070] S103, input the shallow features into the mask generator to generate a domain-invariant mask and a domain-specific mask respectively, and perform channel-wise weighted summation of the domain-invariant mask and the domain-specific mask with the shallow features respectively to obtain the domain-invariant features and the domain-specific features;

[0071] S104, determine the mutual information minimization loss based on the domain-invariant mask and the domain-specific mask, determine the feature reconstruction loss based on the domain-invariant features, the domain-specific features, and the shallow features, determine the triplet loss based on the triplet, and determine the source domain classification loss based on the domain-invariant features and the fault category label; train the feature extractor, the mask generator, and the classifier based on the joint objective function containing the source domain classification loss, the triplet loss, the mutual information minimization loss, and the feature reconstruction loss, wherein the training process does not use target domain device data;

[0072] S105, the mechanical vibration signal to be diagnosed of the target domain device is input into the trained feature extractor and mask generator to obtain the domain-invariant features of the mechanical vibration signal to be diagnosed, and the classifier outputs the fault category based on the domain-invariant features.

[0073] The functions of each module in the model provided in this embodiment of the invention are as follows:

[0074] Shallow feature extractors are used to extract shallow convolutional features rich in temporal details from input source domain samples, avoiding excessive abstraction of device-specific information by deep networks;

[0075] The mask generator generates domain-invariant and domain-specific masks respectively, and decouples shallow features into domain-invariant and domain-specific features after channel-wise weighting.

[0076] A classifier, connected only to domain-invariant features, is used to classify source domain samples based on the decoupled domain-invariant features and predict fault categories.

[0077] The shallow feature extractor provided in this embodiment of the invention is constructed using the early convolutional layers of a one-dimensional deep shrinking residual network (1D-DRSN) to extract shallow features rich in temporal details, thereby avoiding over-encoding of device-specific information by deep networks.

[0078] In this feature space, an anchor-positive-negative sample triplet sampling strategy is designed, and an adaptive boundary triplet loss is introduced to force the clustering of similar fault samples and the separation of dissimilar samples, thereby enhancing the discriminative ability of the features.

[0079] The mask generator provided in this embodiment of the invention receives shallow features output by a shallow feature extractor and generates a domain-invariant mask and a domain-specific mask respectively; the domain-invariant mask is weighted channel-by-channel with the shallow features to obtain the domain-invariant features, and the domain-specific mask is weighted channel-by-channel with the shallow features to obtain the domain-specific features.

[0080] The mask generator provided in this embodiment of the invention is implemented as follows:

[0081] The input shallow features are processed by convolution and batch normalization, then introduced into a random perturbation layer. A Softmax layer maps the intermediate features to a channel probability distribution, followed by convolution, batch normalization, pooling, and activation. Two sub-networks output domain-invariant and domain-specific masks, respectively. These masks are not pre-defined as strictly complementary across channels, and are weighted across channels with the shallow features to obtain domain-invariant and domain-specific features. .

[0082] The loss terms in the joint objective function provided in S104 of this embodiment are defined as follows:

[0083] The source domain classification loss uses cross-entropy loss to supervise the training of the shallow feature extractor and classifier, enabling them to accurately identify fault categories in the source domain.

[0084] The adaptive triplet loss is used to construct anchor-positive-negative sample triplets in the shallow feature space, forcing the clustering of similar fault samples and the separation of dissimilar samples, thereby enhancing the discriminative ability of the features. Its boundary parameters are dynamically adjusted according to the anchor category.

[0085] To minimize the mutual information loss, a mutual information neural estimation framework is used to estimate the mutual information between a domain-invariant mask and a domain-specific mask, thereby reducing the statistical dependence between the two masks and prompting them to form differentiated channel allocation methods.

[0086] The feature reconstruction loss uses the Frobenius norm to measure the difference between the reconstructed features obtained by combining domain-invariant features and domain-specific features and the original shallow features, so that the reconstructed features approximate the original shallow features.

[0087] like Figure 2 As shown, to address the challenges of highly heterogeneous data distribution from multiple sources in real-world industrial scenarios, where the target device is completely invisible during the training phase, making it difficult for traditional fault diagnosis models to generalize, this invention provides a mask-guided feature decoupling multi-source domain generalization cross-device fault diagnosis method (MGFD), comprising the following steps:

[0088] Step 1, Multi-source domain data acquisition: Rolling bearing operation data from multiple different devices (source domains) are collected using vibration sensors. Each source domain device has a complete fault category label. Meanwhile, the target domain devices are not involved at all during the training phase and have no data (including unlabeled data). In this embodiment, the source domain devices represent different operating conditions from multiple publicly available bearing datasets.

[0089] Step 2, Data Preprocessing and Augmentation: The collected vibration signals from each source domain are sliced ​​using the overlapping sliding window method to form multiple samples for model input, thus constructing a labeled source domain sample set. All samples were standardized to eliminate dimensional differences introduced by different devices, sensors, or acquisition environments. Simultaneously, to enhance model robustness, Gaussian perturbation noise was added to the original vibration signals, with the perturbation intensity proportional to the signal standard deviation. The settings simulate random disturbances in an industrial field environment.

[0090] Step 3, Construction of a cross-device diagnostic model based on deep networks: The diagnostic model in this embodiment consists of the following three core parts: shallow feature extractor Mask generator Classifier :

[0091] Shallow feature extractor The initial convolutional layers of a one-dimensional deep shrinking residual network (1D-DRSN) are used (specifically, the first three residual stages, with 512 output channels) to receive source domain samples and map them to a shallow feature space rich in temporal details. This avoids excessive abstraction of device-specific information in deep networks.

[0092] Mask generator It consists of a deep neural network that generates domain-invariant masks. and domain-specific mask The input is shallow features. The process sequentially involves 1×1 convolutional dimensionality reduction, a random perturbation layer, softmax probabilistic modeling, reconvolution, and global max pooling. Finally, a sigmoid activation is applied to output a channel mask vector with values ​​between [0,1]. , Achieve feature decoupling.

[0093] Classifier It consists of a two-layer fully connected network (256→512→256→number of classes) and a Softmax activation function, and only applies to domain-invariant features. Classify and predict the type of fault.

[0094] Step 4, Feature discrimination enhancement: In the shallow feature space, design an anchor-positive-negative sample triplet sampling strategy and introduce adaptive boundary triplet loss. For each training batch, for each anchor sample, select the positive sample with the furthest distance among similar faults and the negative sample with the closest distance among dissimilar faults, and set the boundary values. The loss is dynamically adjusted based on the anchor point category. This loss forces similar faulty samples to cluster in the feature space and separates dissimilar samples, thereby enhancing the discriminative power of the features.

[0095] Step 5, Mask-Guided Feature Decoupling and Constraints: The domain-invariant mask and domain-specific mask generated in Step 3 are applied to the shallow features respectively to obtain domain-invariant features and domain-specific features. The mutual information between the domain-invariant mask and the domain-specific mask is estimated using the MINE (Mutual Information Neural Estimation) framework, and the statistical dependency between the two masks is reduced by minimizing the loss through mutual information. Simultaneously, the domain-invariant features and domain-specific features are combined to obtain reconstructed features. The Frobenius norm is used to calculate the difference between the reconstructed features and the original shallow features, making the reconstructed features approximate the original shallow features to reduce information loss during feature decoupling. ;

[0096] Step 6, Joint Model Optimization and Training: Input the sample sets from each source domain into the model constructed in Step 3 for end-to-end joint training. The goal of training is to minimize a joint loss function that integrates four loss terms to collaboratively optimize the model parameters.

[0097] The overall loss function is:

[0098] ,

[0099] in, For source domain classification loss, cross-entropy loss is used, which only applies to domain-invariant features. This ensures that the model has high classification accuracy on labeled data from multiple sources.

[0100] hyperparameters set to , , (Determined via grid search). By minimizing the joint objective function through backpropagation and the Adam optimizer (learning rate 0.001, batch size 64), the model learns a feature space that is both device-independent (domain-invariant) and highly discriminative, while explicitly separating device-specific interference.

[0101] Step 7, Model Testing and Zero-Shot Fault Diagnosis: The trained diagnostic model is used to diagnose faults on completely unseen samples in the target domain. Since the model has never encountered any target domain data during training, and has only learned how to decouple domain-invariant fault features through multi-source domain training, the shallow feature extractor in the model will perform well when a test sample from the target domain is input. Feature extraction, mask generator Automatically separate domain-invariant features Classifier It directly outputs the corresponding fault category. The entire process requires no target domain data adaptation or fine-tuning, thus achieving zero-shot cross-device fault diagnosis.

[0102] like Figure 3 As shown, this embodiment of the invention provides a multi-source domain generalized cross-device fault diagnosis system with mask-guided feature decoupling, comprising:

[0103] Multi-source domain data acquisition module, used to acquire tagged mechanical vibration signals from multiple devices with different source domains;

[0104] The data preprocessing and enhancement module is used to slice, standardize, and add Gaussian perturbation noise to the vibration signal to form a source domain enhanced sample set;

[0105] The shallow feature extraction and triplet enhancement module is used to construct a one-dimensional deep shrinkage residual network to extract shallow features and calculate the adaptive triplet loss.

[0106] The mask generation and feature decoupling module is used to generate domain-invariant masks and domain-specific masks, decoupling shallow features into domain-invariant features and domain-specific features.

[0107] The mutual information minimization module is used to estimate the mutual information between a domain-invariant mask and a domain-specific mask, and to generate a mutual information minimization loss to reduce the statistical dependence of the two masks.

[0108] The feature reconstruction module is used to calculate the reconstruction loss between the original shallow features and the reconstructed features.

[0109] The classification module, connected only to domain-invariant features, is used to predict fault categories;

[0110] The model training module is used to input multiple source domain sample sets into the model and perform end-to-end training on the model by jointly optimizing the source domain classification loss, adaptive triplet loss, mutual information minimization loss and feature reconstruction loss.

[0111] The fault diagnosis module receives the trained model, processes the target domain test data that has never been seen before, and outputs the final fault category.

[0112] Example 1

[0113] This invention provides a specific application of the mask-guided feature decoupling multi-source domain generalized cross-device fault diagnosis method, as detailed below:

[0114] Step 1, Data Acquisition: The operating status of the equipment under different working conditions is monitored using sensors, and vibration signals from the rolling bearings are collected to construct a multi-source domain generalization task. This embodiment uses four publicly available bearing datasets for verification:

[0115] (1) Case Western Reserve University (CWRU) bearing dataset: Vibration signals were collected under various loads and speeds by pre-setting single faults of different sizes on a 1.5kW motor test bench using EDM technology.

[0116] (2) Jiangnan University (JNU) bearing dataset: On the centrifugal fan system test platform, for two types of bearings, N205 and NU205, signals were collected at different speeds by pre-setting faults through wire cutting.

[0117] (3) Southeast University (SEU) bearing dataset: On the transmission system dynamic simulator (DDS) test bench, various bearing status data, including single and compound faults, are collected by adjusting the speed and load.

[0118] (4) Shandong University of Science and Technology (SDUST) bearing dataset: Vibration signals were collected under different speed and load conditions using a 6205 deep groove ball bearing on a rotating mechanical transmission component fault implantation test bench.

[0119] Step 2, Data Preprocessing and Generalization Task Construction: The original signals from each dataset are sliced ​​using overlapping sliding windows (window size 3072, step size 512) to construct samples. To evaluate model performance, this embodiment selects two typical operating conditions from each of the four datasets, forming eight sets of operating conditions with significant differences in equipment and conditions (A1, A2, B1, B2, C1, C2, D1, D2), which are then uniformly classified into four fault categories: normal, inner race fault, outer race fault, and rolling element fault. By combining operating conditions from different datasets as a multi-source domain training set and using the remaining datasets as an invisible target domain, a dual-source domain generalization task (24 sets) and a tri-source domain generalization task (64 sets) are constructed.

[0120] Step 3, Model Construction: In accordance with the technical solution described in this invention, a deep network model (MGFD) is constructed, consisting of a shallow feature extractor (based on 1D-DRSN), a mask generator (dual branch), a classifier, and a mutual information estimation network.

[0121] Step 4, Model Training: Input the preprocessed sample sets from multiple source domains into the model, and perform end-to-end joint training by minimizing the joint objective function (classification loss + triplet loss + mutual information loss + reconstruction loss). Target domain data is not involved in the training at all.

[0122] Step 5, Model Testing: Input the test data that was not seen during training into each target domain into the trained model. The model directly outputs the predicted fault category to evaluate the zero-shot generalization performance.

[0123] Step 6, Performance Verification and Comparative Analysis:

[0124] To verify the effectiveness of the proposed method (MGFD), this study comprehensively compared it with a benchmark model and three advanced methods. The comparison methods are as follows:

[0125] ResNet: A baseline model that does not contain any transfer or generalization modules, is trained using only multi-source domain data, and is used to evaluate the generalization ability of direct learning.

[0126] MMD-DANN: A domain adaptation method that combines adversarial learning and multi-kernel maximum mean difference (target domain data is required; here, only source domain training is used for fair comparison).

[0127] MI-DSFL: A mutual information-guided two-branch feature decoupling method for domain generalization fault diagnosis.

[0128] SCRL: A semantically consistent representation learning framework for fault diagnosis of domain generalization under class imbalance.

[0129] All comparison methods used the same backbone network (the early layer of 1D-DRSN) as the feature extractor as the present invention, and were tested on a unified hardware and software platform with optimized hyperparameters to ensure the fairness of the comparison.

[0130] Experimental results:

[0131] Starting with the core evaluation metric of diagnostic accuracy, we compared the models mentioned above on multiple cross-device generalization tasks. The results are as follows: Figures 4-8 As shown.

[0132] Dual-source domain generalization results ( Figure 4 The present invention (MGFD) achieved an average accuracy of 43.92% across all 24 cross-device tasks, significantly higher than SCRL (37.16%) and MI-DSFL (36.32%), while the benchmarks ResNet and MMD-DANN were only 23.20% and 23.66%, respectively. In some difficult tasks (such as B1, C1→A1), the present invention achieved an accuracy of 47.37%, with a lead of more than 10 percentage points.

[0133] Three-source domain generalization results ( Figure 5-8 When trained using three source domains and with a completely unknown device as the target domain, the average accuracy of this invention on 64 tasks is as follows: 51.42% with SDUST as the target domain, 53.57% with CWRU as the target domain, 51.70% with JNU as the target domain, and 52.36% with SEU as the target domain. All results outperform the comparative methods and exhibit minimal performance fluctuations, verifying the robustness of this invention in extreme generalization scenarios.

[0134] In summary, the model of this invention performs excellently in terms of diagnostic accuracy, generalization stability, and feature learning quality. It provides an efficient, robust, and highly valuable method for solving the problem of zero-sample cross-device fault diagnosis in industrial practice where there is no data for the target equipment.

[0135] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-source domain generalized cross-device fault diagnosis method with mask-guided feature decoupling, characterized in that, Includes the following steps: Step 1: Obtain mechanical vibration signals and corresponding fault category labels from at least two source domain devices; perform sample division and standardization on the mechanical vibration signals to obtain multiple labeled source domain sample sets. Step 2: Input the source domain samples into the feature extractor to extract shallow features, and construct a triplet consisting of anchor samples, positive samples, and negative samples based on the fault category labels of the source domain samples. Step 3: Input the shallow features into the mask generator to generate a domain-invariant mask and a domain-specific mask respectively. Then, perform channel-wise weighted summation of the domain-invariant mask and the domain-specific mask with the shallow features to obtain the domain-invariant features and the domain-specific features respectively. Step 4: Determine the mutual information minimization loss based on the domain-invariant mask and the domain-specific mask; determine the feature reconstruction loss based on the domain-invariant features, domain-specific features, and shallow features; determine the triplet loss based on the triplet; and determine the source domain classification loss based on the domain-invariant features and fault category labels. Train the feature extractor, mask generator, and classifier using a joint objective function that includes the source domain classification loss, triplet loss, mutual information minimization loss, and feature reconstruction loss, wherein the training process does not use target domain device data. Step 5: Input the mechanical vibration signal to be diagnosed from the target domain device into the trained feature extractor and mask generator to obtain the domain-invariant features of the mechanical vibration signal to be diagnosed, and the classifier outputs the fault category based on the domain-invariant features.

2. The method as described in claim 1, characterized in that, The functions of each module in the model are as follows: Shallow feature extractors are used to extract shallow convolutional features rich in temporal details from input source domain samples, avoiding excessive abstraction of device-specific information by deep networks; The mask generator generates domain-invariant and domain-specific masks respectively, and decouples shallow features into domain-invariant and domain-specific features after channel-wise weighting. A classifier, connected only to domain-invariant features, is used to classify source domain samples based on the decoupled domain-invariant features and predict fault categories.

3. The method as described in claim 2, characterized in that, The shallow feature extractor is constructed using the early convolutional layers of a one-dimensional deep shrinking residual network (1D-DRSN) to extract shallow features rich in temporal details, thus avoiding over-encoding of device-specific information by deep networks. In this feature space, an anchor-positive-negative sample triplet sampling strategy is designed, and an adaptive boundary triplet loss is introduced to force the clustering of similar fault samples and the separation of dissimilar samples, thereby enhancing the discriminative ability of the features.

4. The method as described in claim 2, characterized in that, The mask generator receives the shallow features output by the shallow feature extractor and generates a domain-invariant mask and a domain-specific mask respectively. The domain-invariant mask is weighted channel-by-channel with the shallow features to obtain the domain-invariant features, and the domain-specific mask is weighted channel-by-channel with the shallow features to obtain the domain-specific features.

5. The method as described in claim 4, characterized in that, The mask generator is implemented as follows: The input shallow features are processed by convolution and batch normalization, and then a random perturbation layer is introduced. The intermediate features are mapped to channel probability distributions through a Softmax layer, followed by convolution, batch normalization, pooling, and activation processing. The two sub-networks output domain-invariant masks and domain-specific masks, respectively. The two masks are not pre-defined to be strictly complementary in each channel, and are weighted with the shallow features in each channel to obtain domain-invariant features and domain-specific features.

6. The method as described in claim 1, characterized in that, The loss terms in the joint objective function described in step 4 are defined as follows: The source domain classification loss uses cross-entropy loss to supervise the training of the shallow feature extractor and classifier, enabling them to accurately identify fault categories in the source domain. The adaptive triplet loss is used to construct anchor-positive-negative sample triplets in the shallow feature space, forcing the clustering of similar fault samples and the separation of dissimilar samples, thereby enhancing the discriminative ability of the features. Its boundary parameters are dynamically adjusted according to the anchor category. To minimize the mutual information loss, a mutual information neural estimation framework is used to estimate the mutual information between a domain-invariant mask and a domain-specific mask, thereby reducing the statistical dependence between the two masks and prompting them to form differentiated channel allocation methods. The feature reconstruction loss uses the Frobenius norm to measure the difference between the reconstructed features obtained by combining domain-invariant features and domain-specific features and the original shallow features, so that the reconstructed features approximate the original shallow features.

7. A multi-source domain generalized cross-device fault diagnosis system implementing the method of any one of claims 1-6, characterized in that, include: Multi-source domain data acquisition module, used to acquire tagged mechanical vibration signals from multiple devices with different source domains; The data preprocessing and enhancement module is used to slice, standardize, and add Gaussian perturbation noise to the vibration signal to form a source domain enhanced sample set; The shallow feature extraction and triplet enhancement module is used to construct a one-dimensional deep shrinkage residual network to extract shallow features and calculate the adaptive triplet loss. The mask generation and feature decoupling module is used to generate domain-invariant masks and domain-specific masks, decoupling shallow features into domain-invariant features and domain-specific features. The mutual information minimization module is used to estimate the mutual information between a domain-invariant mask and a domain-specific mask, and to generate a mutual information minimization loss to reduce the statistical dependence of the two masks. The feature reconstruction module is used to calculate the reconstruction loss between the original shallow features and the reconstructed features. The classification module, connected only to domain-invariant features, is used to predict fault categories; The model training module is used to input multiple source domain sample sets into the model and perform end-to-end training on the model by jointly optimizing the source domain classification loss, adaptive triplet loss, mutual information minimization loss and feature reconstruction loss. The fault diagnosis module receives the trained model, processes the target domain test data that has never been seen before, and outputs the final fault category.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the mask-guided feature decoupling multi-source domain generalization cross-device fault diagnosis method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer program is stored therein, and when executed by a processor, the processor performs the steps of the multi-source domain generalized cross-device fault diagnosis method as described in any one of claims 1-6.

10. An information data processing terminal, characterized in that, The information data processing terminal includes the multi-source domain generalized cross-device fault diagnosis system as described in claim 7, which features mask guidance and decoupling.

Images