Industrial equipment adaptive fault diagnosis method and system

Abstract

The invention provides a self-adaptive fault diagnosis method and system for industrial equipment, and belongs to the field of intelligent fault diagnosis, and the method comprises the steps: obtaining equipment fault vibration data of an industrial scene; dividing the equipment fault vibration data into a source domain sample and a target domain sample; and processing the equipment fault vibration data based on the universal field adaptive model to obtain a fault classification result of the universal field, and performing fault diagnosis on the industrial equipment based on the fault classification result. Cross-domain knowledge migration and diagnosis optimization in a complex tag space are achieved by suppressing interference of private classes of a source domain and dynamically discovering unknown classes of a target domain.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis technology, specifically relating to an adaptive fault diagnosis method and system for industrial equipment. Background Technology

[0002] Deep learning technology has been widely applied in fields such as image recognition and fault diagnosis. Deep neural networks can automatically learn complex features in data, thereby improving the accuracy of recognition and diagnosis. Against the backdrop of industrial transformation towards intelligence, intelligent fault diagnosis methods based on deep neural networks (DNNs) are widely used. By constructing multi-layered feature representation networks, complex and non-linear feature patterns are automatically extracted, improving the diagnostic system's accuracy in identifying fault types and its generalization ability. However, the success of deep learning technology largely depends on the availability of a large amount of labeled data, which is time-consuming and costly to collect. In practical applications, data from different domains often exhibit significant distribution differences (i.e., domain shift), making it difficult for models trained in one domain to directly generalize to a completely new and unlabeled domain. For fault diagnosis of industrial equipment, different operating conditions, equipment status, and environmental factors can all cause changes in data distribution, affecting the diagnostic results. To address this issue, unsupervised domain adaptation (UDA) has been proposed. By narrowing the feature differences between the source domain (rich in labeled data) and the target domain (scarce in labeled data), it achieves cross-domain knowledge transfer and has been successfully applied in multiple fields such as image recognition, speech processing, and fault diagnosis.

[0003] Traditional unsupervised domain adaptation methods mostly assume that the source and target domains share the same label space, a concept known as close-set domain adaptation (CDA). However, this assumption is often difficult to hold in real-world scenarios, especially in industrial settings where the target domain may contain new categories not seen in the source domain, such as novel fault types or equipment anomalies. To address these challenges, more general domain adaptation frameworks have been explored, such as Partial Domain Adaptation (PDA) and Open-set Domain Adaptation (OSDA). These methods attempt to introduce class selection mechanisms, open class detection mechanisms, or uncertainty modeling mechanisms to improve the model's adaptability in partially overlapping label spaces. However, these methods are often designed for specific scenarios, lacking uniformity and generalization, and are not highly applicable to the highly uncertain, complex, and varied distributions and class differences encountered in real-world industrial settings. Summary of the Invention

[0004] To address the shortcomings of existing general domain adaptation frameworks in fault diagnosis, this invention provides an adaptive fault diagnosis method and system for industrial equipment.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An adaptive fault diagnosis method for industrial equipment includes the following steps: Acquire equipment fault vibration data in an industrial setting; divide the equipment fault vibration data into source domain samples and target domain samples; The equipment fault vibration data is processed using a general-domain adaptive model to obtain a general-domain fault classification result. Fault diagnosis of industrial equipment is then performed based on this classification result. In the general-domain adaptive model, the source domain is weighted, and a feature extractor is used to represent the weighted source domain samples and target domain samples. Based on these feature representations, a discriminator distinguishes the feature distributions of the source and target domains, and the mean value of the features is taken by category to obtain the prototype vector for each known class. Similarity calculation is performed between the target domain samples and the prototype vectors of each known class. Based on a similarity threshold, the sample with the lowest similarity to the source domain prototype is selected as the initial unknown class prototype. Based on the nearest neighbor rule, samples in the target domain that are the nearest neighbors to the initial unknown class prototype are assigned to the current unknown class. This process is iteratively repeated, selecting all target domain samples as new unknown class prototypes, to obtain a preliminary clustering result for the target domain. The preliminary clustering result is corrected by calculating the sum of squared clustering errors, and the general-domain adaptive classification result is output.

[0006] Preferably, the prototype vector of each known class is specifically: ; in, Let n represent the prototype of the k-th class. k This indicates the size of the number of samples included in the category.

[0007] Preferably, it also includes providing a source domain sample weighting mechanism based on category confidence awareness to reconstruct the preliminary clustering results. Specifically, it uses the prediction confidence of the target domain to the source category, where the confidence is the desired weight, to dynamically adjust the weighting weights during weighted reconstruction, thereby suppressing the influence of private categories in the source domain during domain adversarial learning. This includes the following steps: The average predicted probability of the target domain sample for the source class is calculated using the following formula: ; in, This refers to the softmax function. It is a predictive classifier. ∈{1,..., } is used to represent the source domain category index. Represents the number of samples in the target domain; Based on the source sample label y si Mapping class weights: ; Among them, y si For source sample x si The true label; By incorporating sample weights into the adversarial loss function, feature alignment of common class features in the source domain is achieved.

[0008] Preferably, incorporating sample weights into the adversarial loss function specifically involves: ; Among them, L d (.) represents the domain label prediction loss. and These represent the real domain labels of the source and target domains, respectively. Indicates the discriminator, Indicates feature extractor, Indicates the number of samples in the target domain. Indicates the number of samples in the source domain. Indicates the loss weight. This represents the i-th sample in the source domain. This indicates the number of samples in the source domain.

[0009] Preferably, the method further includes mitigating the excessive aggregation of unknown class features in the target domain by performing focus diversity processing on the target domain. Specifically, the focus diversity processing involves L2 normalization of the target domain samples and then calculating the cosine similarity of the sample pairs; and using a focus weight mechanism to set penalty weights for samples with high similarity.

[0010] Preferably, the general domain adaptive model further includes a classifier, which is used to predict the class label of a sample based on the feature vector generated by the feature extractor; for source domain samples, the classifier is trained based on the labeled tags to learn the mapping relationship between features and known categories; for target domain samples, the classifier predicts the known category to which the target domain sample belongs based on the mapping relationship learned from the source domain.

[0011] This invention also provides an adaptive fault diagnosis system for industrial equipment, specifically comprising: The data acquisition module is used to acquire equipment fault vibration data in industrial scenarios; and to divide the equipment fault vibration data into source domain samples and target domain samples.

[0012] The classification processing module is used to process the equipment fault vibration data based on a general domain adaptive model to obtain a general domain fault classification result, and to diagnose industrial equipment faults based on the fault classification result. In the general domain adaptive model, the source domain is weighted, and a feature extractor is used to represent the weighted source domain samples and target domain samples as features. Based on the feature representation, a discriminator distinguishes the feature distributions of the source domain and the target domain, and the feature mean is taken according to the category to obtain the prototype vector of each known class. The similarity between the target domain samples and the prototype vectors of each known class is calculated. Based on the similarity threshold, the sample with the lowest similarity to the source domain prototype is selected as the initial unknown class prototype. Based on the nearest neighbor rule, the samples in the target domain that are the nearest neighbors to the initial unknown class prototype are assigned to the current unknown class. All target domain samples are iteratively selected as new unknown class prototypes, and the sample allocation process is repeated to obtain the preliminary clustering result of the target domain. The preliminary clustering result is corrected by calculating the sum of squared clustering errors of the preliminary clustering result, and the general domain adaptive classification result is output.

[0013] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the adaptive fault diagnosis method for industrial equipment.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute the steps described in the adaptive fault diagnosis method for industrial equipment.

[0015] The adaptive fault diagnosis method for industrial equipment provided by this invention has the following beneficial effects: This invention measures the similarity between source and target domain samples, prioritizing the allocation of target domain samples to the nearest source domain prototype. This strengthens the alignment of common class features, suppresses interference from private classes, reduces reliance on large-scale labeled data, and saves significant manpower and time. A mutual nearest neighbor comparison learning mechanism is designed to guide the gradual formation of clustering structures for unknown categories in the target domain, achieving fine-grained partitioning and improving the ability to distinguish multiple potential unknown classes in the target domain. This results in a complete, general-domain adaptive fault diagnosis system, enhancing the ability and robustness to identify unknown faults under complex, multi-source, and variable operating conditions. Attached Figure Description

[0016] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a diagram of an adaptive fault diagnosis method for industrial equipment according to the present invention.

[0018] Figure 2 This is a schematic diagram of the progressive recognition process in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram illustrating focus diversity processing in an embodiment of the present invention.

[0020] Figure 4 This is a feature visualization comparison of migration task A3 in an embodiment of the present invention. Figure 4 (a) shows the feature visualization effect of the DANN algorithm; Figure 4 (b) shows the feature visualization effect of the OSBP algorithm; Figure 4 (c) shows the feature visualization effect of the UAN algorithm; Figure 4 (d) represents the feature visualization effect of the OVANet algorithm; Figure 4 (e) represents the feature visualization effect of the IUAN algorithm; Figure 4 (f) represents the visual effect of the features of the method proposed in this invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0022] Example This invention provides an adaptive fault diagnosis method for industrial equipment, such as... Figure 1 As shown, it specifically includes: Feature alignment; known class transfer, unknown class modeling; structural optimization; knowledge transfer and diagnostic optimization.

[0023] Step 1: (1) A weighted adversarial training mechanism is used to achieve cross-domain feature alignment. In this process, the feature extractor and discriminator are optimized together with the help of adversarial learning strategy to obtain feature representations with domain invariance. During training, the discriminator focuses on distinguishing the feature distributions of the source domain and the target domain, while the feature extractor generates a feature space that is difficult to distinguish with the help of the adversarial mechanism. When there are unknown class samples in the target domain, traditional adversarial training may cause these samples to be incorrectly classified into known classes in the source domain, resulting in negative transfer. To deal with this problem, a similarity-weighted optimization process is designed. This process relies on calculating the similarity measure between the target sample and the source domain prototype. The calculation of the cosine similarity between the target sample and the source prototype is as follows:

[0024] ; in, It is the first in the source domain The prototype of the class sample, Let j represent the j-th sample in the target domain.

[0025] Prioritizing the assignment of samples to the nearest source domain prototype effectively suppresses the interference of unknown samples on model performance. In detail, the weighted adversarial loss function is calculated as follows: ; in, Indicates the domain label prediction loss. and These represent the real domain labels of the source and target domains, respectively. Indicates feature extractor, Represents the i-th sample in the source domain. Indicates the number of samples in the source domain. Indicates discriminator, Indicates loss weight, Represents the number of samples in the target domain, n s This represents the number of samples in the source domain.

[0026] (2) By using weighted adversarial methods, the negative impact of unknown samples on the process is suppressed while aligning the feature distributions of the source and target domains. However, relying solely on global feature alignment strategies can easily lead to confusion between known categories and may cause unknown category samples to be incorrectly classified into known categories. To address this limitation, a progressive recognition process is designed, such as... Figure 2 As shown, this approach achieves refined category alignment between known category samples in the source and target domains, and enables more accurate classification of unknown samples in the target domain. Since the target domain samples lack labeled information, a transfer learning strategy based on source domain sample label knowledge is employed to guide the exploration of target domain samples in the category space. This is specifically achieved by calculating the source domain category prototype, the mathematical formula of which is as follows:

[0027] ; in, It is the first in the source domain The prototype of the class sample, Used to indicate the size of the number of samples contained in that category. Represents the i-th sample in the source domain. Indicates feature extractor, Indicates the nth sample.

[0028] To ensure the robustness of the local feature alignment process, a nearest neighbor matching strategy is implemented between the source domain prototype and the target domain samples to handle samples of known categories. Based on a contrastive learning framework, to improve the similarity measurement between nearest neighbor sample pairs and suppress their similarity to other samples, the contrastive learning loss function constructed in this way can be expressed in the following mathematical form:

[0029] ; in, For temperature control parameters, , It is the set of mutually nearest neighbor matching pairs (here, the degree of mutual proximity in the source domain and the sample domain is calculated first, and the loss function is only calculated for mutually nearest neighbor matching pairs). This represents the features extracted from the target domain samples by the feature extractor.

[0030] When extracting features from samples in the target domain, the feature vector z is generated by the feature extractor. Based on the similarity measure between the source domain prototype and the target domain samples, the nearest neighbor matching mechanism can effectively identify the nearest sample pairs in the feature space, thereby achieving more accurate cross-domain feature alignment and reducing inter-class confusion. When there is a significant difference in class distribution between the source and target domains, the representation of unknown class samples in the feature space is often significantly different from the known classes in the source domain. Therefore, selecting the target domain sample with the lowest similarity to the source domain prototype as the initial prototype of the unknown class can provide a clearly discriminative initial benchmark for subsequent unknown class feature learning and fine-grained classification tasks. The mathematical expression of this initial prototype can be expressed as follows:

[0031] ; Here, sim(.) represents the cosine similarity function. After initializing the unknown class prototypes, a nearest neighbor contrastive learning strategy is used to classify the samples. Specifically, by selecting samples in the target domain that form nearest neighbor relationships with the initialized prototypes, these samples have the closest distribution in the feature space to the prototypes and have a high probability of belonging to the same category. Based on this, a contrastive learning mechanism is designed to promote a closer clustering distribution of related samples in the feature space. The corresponding loss function can be expressed as:

[0032] ; This method employs a progressive identification mechanism that selects the target sample with the greatest difference from the source domain features as the new prototype. It uses an iterative approach to construct the prototype representation of the unknown category and combines a nearest neighbor comparison learning algorithm to achieve automatic sample allocation until all unknown samples are effectively classified, ultimately achieving a fine-grained classification goal. This approach ensures the model's recognition performance for known categories in the target domain and also provides it with robustness to dynamically adapt to the evolution of unknown categories, thus improving the adaptability of the classification system in open environments.

[0033] (3) Verification Mechanism: To accurately infer the number of unknown categories in the target domain and deeply explore its potential category structure, a verification mechanism is introduced. This verification mechanism performs cluster optimization based on the preliminary classification results output by the progressive identification process. By effectively integrating similar categories to eliminate redundant classifications, it improves the model's discrimination accuracy for unknown category samples in the target domain. In the progressive identification stage, the model uses an iterative approach to gradually complete the identification and classification of unknown category samples. However, since the unknown categories in the target domain may have complex distribution characteristics, direct classification can easily lead to an increase in the number of categories and cause overclassing. The verification mechanism adopts a post-processing strategy, using system evaluation and cluster fusion to optimize the identified categories, and reasonably merges categories with similar features or spatial overlap. While ensuring appropriate classification granularity, it maintains the theoretical rationality of the number of categories. In specific implementation, the verification mechanism first performs similarity analysis on the initial categories generated by progressive identification, reconstructs the cluster structure based on the spatial distribution of sample features, and performs merging operations on categories with proximity relationships in the feature space. This effectively controls the total number of categories and improves intra-class homogeneity. This optimization process improves the overall classification performance and enhances the distinguishability between categories. In terms of technical implementation, the verification mechanism first evaluates the current classification state by calculating the sum of squared errors of each cluster, and its mathematical expression is as follows:

[0034] ; In the algorithm implementation, the `nclusters` parameter describes the total number of clusters in the current clustering partition. When merging and evaluating any two cluster subsets, the system first constructs candidate partitioning schemes containing `nclusters-1` categories and calculates their corresponding sum of squared errors (SSEnew). Based on the comparative analysis of the sum of squared errors before and after merging, the algorithm dynamically decides whether to perform the cluster merging operation, thereby achieving adaptive optimization of the category structure. The verification mechanism, through iterative evaluation and adjustment, effectively avoids classification redundancy caused by overfitting while ensuring significant distinction between categories. This dynamic optimization strategy improves the model's category recognition ability in open scenarios, enhances the generalization performance of the classification system, and ensures the rationality and effectiveness of the classification results.

[0035] (4) Reconstruction Mechanism: To effectively retain the important information needed to distinguish between unknown and known categories, a reconstruction mechanism is introduced. This mechanism relies on feature space mapping and reconstruction to effectively preserve the discriminative features of both unknown and known categories. The core idea of ​​the reconstruction mechanism is to project the target domain samples into a new representation space. During this process, it focuses on extracting highly discriminative feature representations, thereby simultaneously improving the model's accuracy in recognizing known categories and its ability to distinguish unknown categories. In terms of implementation, the reconstruction mechanism utilizes feature mapping and reconstruction to ensure that the reconstructed feature representations fully retain key category discrimination information, especially those specific features that contribute significantly to the detection of unknown categories. This design improves the feature recognition of target domain samples and systematically optimizes the model's performance in open recognition tasks. To enhance the relevance and effectiveness of the reconstruction process, a weighted reconstruction loss function is introduced. Its key feature is the dynamic adjustment of the weight allocation strategy for the reconstruction error based on the critical differences between samples and the classification task, allowing the model to focus on sample features that have a crucial impact on classification decisions. The mathematical definition of this weighted reconstruction loss function is as follows:

[0036] ; Among them, x j Let the j-th target domain sample be denoted as . Introduce the reconstruction weight parameter w j This weighted optimization mechanism, which uses weights to represent the criticality of each sample, allows the model to adaptively focus on samples with larger reconstruction errors or obvious discriminative features. This strategy improves the classification performance of samples with known categories and effectively maintains the potential structural features of samples with unknown categories, thereby enhancing the model's ability to distinguish unknown categories in the target domain.

[0037] Step Two: Source Domain Weighting Mechanism. An adversarial training strategy that treats all source domain samples equally can cause private categories in the source domain to interfere with the cross-domain alignment of common category samples. Forced alignment of private category samples distorts the feature distribution of common categories, leading to a decrease in model performance. To mitigate this problem, a source domain weighting mechanism based on category confidence is introduced. This mechanism uses the prediction confidence of the target domain for the source category as the desired weight, dynamically adjusting the weights of source domain samples in adversarial training to suppress the influence of private categories in domain adversarial learning. First, the average prediction probability of the target domain samples for the source category must be calculated:

[0038] ; Where σ refers to the softmax function, and C is the predictor classifier. ∈{1,..., } is used to represent the source domain category index, nt This represents the number of samples in the target domain. The predicted probability of a target domain sample in a private category of the source domain is generally low. The average predicted probability reflects the probability that a category in the original domain is a common category. After normalization, it is calculated based on the source sample label y. si To map category confidence: ; Where y si For source sample x si The true labels are used to convert category-level confidence scores into sample-level weights. The adversarial loss function handled by this mechanism can be expressed as:

[0039] ; This mechanism allows the model to enhance feature alignment between high-confidence categories in the target domain while suppressing the negative effects of source domain private categories on domain adversarial training.

[0040] Step 3: Focus Diversity Handling. In domain adaptation tasks, besides identifying unknown categories, the model must also ensure sufficient diversity in the feature representations of these unknown categories. This prevents all unknown samples from being simply concentrated in a compact feature space region, such as... Figure 3 As shown in the diagram, traditional methods for identifying unknown classes typically focus on distinguishing between unknown and known classes, neglecting the internal structure and diversity of the unknown class. In fact, an unknown class may contain multiple potential new categories. If these are all compressed into the same region of the feature space, the model will be unable to distinguish these potential new categories, reducing the effectiveness of subsequent category discovery and impacting the discriminative power of the feature space.

[0041] The main idea behind focus diversity processing is to promote feature diversification by penalizing high similarity between unknown class samples. Specifically, it involves first performing L2 normalization on the unknown class features, then calculating the cosine similarity matrix between samples, and finally applying a focus weight-based mechanism to assign greater penalty weights to highly similar sample pairs. This loss can be expressed as:

[0042] ; sim(f i f j () refers to the feature vector after normalization. and The loss function uses cosine similarity between samples, with γ being the focus coefficient, which defaults to 2. By employing focus weights, this loss function emphasizes sample pairs that are already highly similar, driving the model to learn to separate them, thereby achieving a diversified distribution of unknown class samples. When the γ value is high, the model will pay more attention to highly similar sample pairs, thus enhancing diversity.

[0043] The method proposed in this invention is further illustrated by the data processing of examples: Data was collected from different platforms, including the bearing fault dataset from Case Western Reserve University's CWRU and the gearbox composite fault dataset from IEEE PHMChallenge.

[0044] The bearing failure dataset from Case Western Reserve University was used, as shown in Table 1. It includes three types of single-point failures: inner ring failure, rolling element failure, and outer ring failure. Vibration signals from the CWRU dataset were collected under three different load conditions by accelerometers mounted on the drive and fan ends of the motor housing, and included four health states: ball failure (BF), inner ring failure (IF), outer ring failure (OF), and healthy (H). Each category in the dataset contains 500 samples, and each sample is 1200 characters long.

[0045] Table 1 CRWU Dataset Settings The second gearbox composite fault dataset. The gearbox equipment configuration includes one input pinion, two idler gears, one output pinion (with 32, 96, 48, and 80 teeth respectively), and six bearings. The vibration acceleration signal was sampled at a frequency of 66.7 kHz. The dataset contains eight different health categories: (1) Normal (N); (2) Gear breakage and eccentricity (CE); (3) Gear eccentricity (E); (4) Gear eccentricity and fracture (EB); (5) Gear breakage, eccentricity, and bearing inner and outer raceway failure (CEF); (6) Gear fracture, bearing inner and outer raceway failure (BF); (7) Bearing inner and outer raceway failure and keyway shear (FS); (8) Bearing ball and outer raceway failure, shaft imbalance (FI). The dataset includes four operating conditions: (30, H), (40, H), (40, L), and (50, H), where the elements represent the input shaft speed (Hz) and load, respectively, H represents high load, and L represents low load.

[0046] The adversarial framework consists of three parts: a feature extractor, a discriminator, and a classifier. The feature extractor contains four convolutional layers and one pooling layer, while the discriminator and classifier each contain two fully connected layers. This deep network is built on the PyTorch 2.0 framework and an NVIDIA GeForce RTX 3090 GPU. The specific structure and parameter settings are shown in Table 2.

[0047] Table 2 Parameters of the Neural Network The vibration signal was segmented using a data overlap partitioning technique, resulting in a sample containing 1024 data points. Increasing the number of samples enhances the robustness of the model.

[0048] To characterize the model's classification performance under different unknown fault types and variations in the number of categories in both domains, several sample sets were selected to combine the source and target domains. In the CWRU dataset, bearing fault data at speeds corresponding to specific fault types were used as the source domain based on the equipment's operating speed (rpm), while bearing fault data at speeds different from the source domain were used as the target domain. Furthermore, the source and target domains had completely different speeds and different label sets within the same task. In the gearbox dataset, equipment operating conditions, including speed and load level, were used to divide the dataset. Gearbox fault data under specific fault type labels and equipment operating conditions were used as the source domain, while gearbox fault data under equipment operating conditions different from the corresponding source domain were selected as the target domain. Detailed information on the open-set domain adaptation tasks on the CWRU and gearbox datasets is provided in Tables 3 and 4, respectively.

[0049] Table 3. Details of the domain adaptation task on the CWRU dataset. Table 4. Details of the domain adaptation task on the Gearbox dataset. The proposed method was compared with the following five domain adaptation methods to verify the superiority of the proposed method: (1) Domain Adversarial Training of Neural Networks (DANN): As a typical closed-domain adaptation method, DANN performs adversarial training through a domain discriminator and a feature extractor, and then learns cross-domain invariant features.

[0050] (2) Universal Domain Adaptive Network (UAN): UAN uses the similarity of samples to weight samples, filters samples of common categories, and rejects samples of unknown categories.

[0051] (3) Universal Domain Adaptation Network (OVANet): OVANet identifies unknown and shared classes in the target domain by learning the minimum inter-class distance in the source domain.

[0052] (4) Backpropagation Open Set Adaptation (OSBP): A commonly used open set DA network, structurally similar to DANN, which uses adversarial training to identify unknown samples.

[0053] (5) Universal Domain Adaptation Based on Pseudo-margins (IUAN): IUAN identifies the common label space through pseudo-margins and is used for sample classification in UDA scenarios.

[0054] Five different domain adaptation methods were compared on two different fault diagnosis datasets, and transfer tasks with nine different label sets were used to better validate the performance of the methods in complex open scenarios. Tables 5 and 6 show the test accuracy of these methods on different transfer tasks.

[0055] Table 5. Classification accuracy (%) of the proposed method and the comparison method in the CWRU task. Table 5 presents the test accuracy of five domain adaptation methods on the CRWU dataset. The experimental results show that the model proposed in this invention achieves an average accuracy of 90.9%, which is better than other domain adaptation methods. Compared with other methods, the model proposed in this invention exhibits higher accuracy in some open-set tasks. This model can effectively identify and detect shared and unknown class features in both the source and target domains, achieving accurate and effective cross-domain alignment. In contrast, DANN, designed as a closed-set domain adaptation method, cannot effectively distinguish private class samples in the domain, causing its cross-domain alignment to be severely interfered with by private class samples in both the source and target domains, resulting in poor performance in complex open-set classification tasks. OSBP, as an open-set domain adaptation method, has a mechanism for identifying unknown classes in the target domain, improving its performance in open-set scenarios. However, because OSBP does not pay attention to private class samples in the source domain, the private samples interfere with the cross-domain alignment of features of common class samples, leading to feature confusion and poor performance in complex open tasks. UAN uses sample similarity for weighted filtering to identify common category samples and reject unknown category samples. However, when the source or target domain contains a large number of unknown categories, the model may over-reject samples, misclassifying some common category samples as unknown classes. OVANet directly identifies unknown classes using a "one-to-many" classifier, but if the number of samples for certain categories in the source domain is small and the corresponding binary classifier is not sufficiently trained, it may incorrectly reject known class samples of that category. IUAN uses a pseudo-residual loss function to increase the discriminative boundary of known category features, and optimizes the residual loss using target domain pseudo-labels, preventing unknown class samples from intruding into the discriminative space of known class samples. It has good recognition accuracy in some open-set domain adaptation tasks. However, the generation of target domain pseudo-labels is overly dependent on the initial performance of the source domain classifier. If the source domain model has generally low confidence in the target domain, it will accumulate and amplify the pseudo-label error, causing the optimization of the residual loss to deviate from the correct direction.

[0056] In contrast, the constructed model, relying on progressive recognition and verification mechanisms, can effectively distinguish between known and unknown class samples in the target domain. Source domain private class suppression effectively mitigates the interference caused by source domain private class samples. Cross-domain alignment is achieved through domain adversarial training, narrowing the domain and class gaps. Focus diversity loss can separate overcompressed unknown class samples in the target domain, preserving subtle differences between different unknown class samples and obtaining a better discrimination space. This model exhibits good feature discrimination capabilities in complex open scenarios, effectively alleviating the problem of sample confusion between domains caused by too many open categories.

[0057] Table 6. Classification accuracy (%) of the proposed method and the comparison method in the Gearbox task. Table 6 presents the test accuracy of five domain adaptation methods on the Gearbox dataset. To comprehensively evaluate model performance, this experiment tested various domain adaptation tasks to simulate the challenges of fault diagnosis in a real industrial environment. The experimental data shows that the model's average diagnostic accuracy reached 91.7%. This result fully verifies the universality of this model as a general domain adaptation method in different tasks. It can handle private category samples in both the source and target domains well and complete the knowledge transfer of common categories. This model combines weighted adversarial training with progressive recognition to improve fault diagnosis performance under different settings. Weighted adversarial learning effectively suppresses the interference of private categories while strengthening the alignment of common category features by dynamically adjusting sample weights. It gradually selects new prototypes by selecting the least dissimilar prototypes and combines them with mutual nearest neighbor contrast learning until fine-grained classification of all unknown samples is completed, achieving accurate detection of new fault types. Focus diversity loss separates overcrowded unknown class samples, obtains a good discrimination space, and improves the recognition accuracy of unknown categories.

[0058] To comprehensively evaluate the impact of the proposed source domain weighting mechanism and focus diversity handling on model performance, ablation experiments were conducted, and their effectiveness was validated on the CWRU and Gearbox datasets. The source domain weighting W... s and focus diversity loss L div These were excluded from the original model to examine their impact on model performance. The average diagnostic accuracy results are shown in Table 7:

[0059] The results show that removing the source domain weighting mechanism reduced the model accuracy by approximately 7%, which fully verifies the crucial role of source domain weighting in improving model performance. Based on the similarity or predicted entropy between the sample and the prototype, different weights can be adaptively assigned to samples, emphasizing information from samples with high confidence, effectively suppressing interference from private class samples, and ensuring accurate cross-domain distribution alignment of common categories. Adding focus diversity processing significantly improved the model's diagnostic performance. Focus diversity processing separates over-aggregated unknown class samples in the target domain, ensuring sufficient discriminative space between different unknown class samples in the feature space, preserving subtle differences, reducing the risk of category confusion, and improving sample discrimination ability. These two processes work together to improve the model's ability to distinguish between known and unknown class samples, while also increasing the classification accuracy for unknown class samples, ultimately resulting in a more accurate classification model.

[0060] Table 7. Statistical results of ablation analysis (%) Taking task A3 as an example, Figure 4 This study presents the distribution of different methods in the feature space for complex transfer tasks where the target domain contains both unknown classes and private classes from the source domain. t-SNE maps features to a two-dimensional space, intuitively showing the feature distribution after domain adaptation, thereby evaluating the performance of domain adaptation and the feature distribution of the target class. Based on these visualizations, the following observations and conclusions can be drawn.

[0061] DANN, as a closed-set domain adaptation method, attempts to align the overall distribution of the source and target domains; however, it fails to consider the challenges posed by class openness. Figure 4 As can be observed in (a), there is a significant overlap among multiple color categories, and the boundary between the unknown target class and the source class is very blurred, resulting in severe feature aliasing. DANN has certain limitations in dealing with open-class scenarios and is prone to negative transfer. OSBP, as an open-set domain adaptive method, introduces a recognition mechanism for unknown classes. Figure 4 As can be seen in (b), the various target features begin to show a clearer separation trend, and the unknown class can also achieve independent clustering to a certain extent. However, since OSBP only focuses on the unknown target class and does not consider the private class of the source domain at the same time, there is still some feature confusion, especially in the boundary region, which is manifested as over-alignment. OVANet, as a general domain adaptive method, identifies unknown class samples through a "one-to-many" classifier and suppresses private class samples through source domain weighting. Figure 4 As can be seen from (d), the boundaries of various target features are relatively clear, and unknown classes can achieve independent clustering, but the overall aggregation effect is not good. Figure 4 (c) and Figure 4 (e) presents the t-SNE visualization results of two general domain adaptation methods, UAN and IUAN. UAN uses sample similarity weighting to filter common categories, enabling various target samples to initially form cluster structures and providing good separability for known categories. However, the distribution of unknown categories is relatively scattered, and there is some overlap with the boundaries of known categories, which still has certain limitations when dealing with complex inter-class relationships. IUAN introduces a pseudo-margin mechanism on the basis of UAN, which effectively improves the ability to distinguish between common and private class samples. From the visualization results, IUAN's feature clusters are more compact and clear, the distribution of unknown class samples is relatively isolated, the boundaries are clear, and it shows stronger discriminative power and domain adaptability.

[0062] The method proposed in this invention is a general domain adaptive strategy that can simultaneously process both unknown target classes and private source domain classes. Figure 4 In (f), each category forms a clear clustering structure with clear boundaries and minimal inter-class interference. This method exhibits superior discriminative ability and transfer adaptability in identifying and isolating different categories, improving the separability of feature distribution and the robustness of domain alignment. Relying on... Figure 4 The t-SNE visualization results clearly show that the proposed method exhibits superior feature discrimination and transfer performance in complex open scenarios, effectively alleviating the confusion caused by complex open categories and verifying the effectiveness of its general domain adaptive capability.

[0063] The present invention also provides an adaptive fault diagnosis system for industrial equipment, comprising: The data acquisition module is used to acquire equipment fault vibration data in industrial scenarios; the equipment fault vibration data is divided into source domain samples and target domain samples.

[0064] The classification processing module is used to process equipment fault vibration data based on a general-domain adaptive model to obtain fault classification results in the general domain. Based on these results, fault diagnosis is performed on industrial equipment. In the general-domain adaptive model, the source domain is weighted, and a feature extractor represents the weighted source and target domain samples. Based on these representations, a discriminator distinguishes the feature distributions of the source and target domains, and the mean value of the features is taken by category to obtain the prototype vector for each known class. Similarity calculation is performed between the target domain samples and the prototype vectors of each known class. Based on a similarity threshold, the sample with the lowest similarity to the source domain prototype is selected as the initial unknown class prototype. Based on the nearest neighbor rule, samples in the target domain that are the nearest neighbors to the initial unknown class prototype are assigned to the current unknown class. This process iteratively selects all target domain samples as new unknown class prototypes, repeating the sample allocation process to obtain preliminary clustering results for the target domain. The preliminary clustering results are corrected by calculating the sum of squared clustering errors, and the general-domain adaptive classification results are output.

[0065] The modules in the aforementioned adaptive fault diagnosis system for industrial equipment can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0066] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of an adaptive fault diagnosis method for industrial equipment. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0067] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the aforementioned method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of an adaptive fault diagnosis method for industrial equipment. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0068] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

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

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

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

[0072] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. An adaptive fault diagnosis method for industrial equipment, characterized by, The method comprises the following steps: obtaining equipment fault vibration data of an industrial scene; dividing the equipment fault vibration data into source domain samples and target domain samples; processing the equipment fault vibration data based on a general domain self-adaptive model to obtain a fault classification result of the general domain, and performing fault diagnosis on an industrial equipment based on the fault classification result; in the general domain self-adaptive model, the source domain is weighted, the weighted source domain samples and target domain samples are represented by a feature extractor, the feature distribution of the source domain and the target domain is distinguished based on the feature representation, the feature mean value of each known class is taken, and a prototype vector of each known class is obtained; similarity calculation is performed on the target domain samples and the prototype vector of each known class, and the sample with the lowest similarity to the source domain prototype is selected as the initial unknown class prototype according to a similarity threshold value, the samples that are the nearest neighbors of each other in the target domain and belong to the initial unknown class prototype are assigned to the current unknown class based on the mutual nearest neighbor rule, all target domain samples are iteratively screened as new unknown class prototypes, and the sample assignment process is repeated to obtain a preliminary clustering result of the target domain; the preliminary clustering result is modified by calculating the sum of squared errors of the clustering result, and an adaptive classification result of the general domain is output.

2. The adaptive fault diagnosis method of industrial equipment according to claim 1, characterized in that, The prototype vector of each known class is specifically: ； wherein, represents a prototype of the kth class, n k represents the number of samples contained in the class.

3. The adaptive fault diagnosis method of industrial equipment according to claim 1, characterized in that, It also includes providing a source domain sample weighting mechanism based on category confidence awareness to weight the preliminary clustering result, specifically by measuring the prediction confidence of the target domain on the source category, the confidence is the weight, the weighting weight during weighting reconstruction is dynamically adjusted, the influence of the source domain private category in domain adversarial learning is suppressed, and the following steps are specifically included: Calculate the average prediction probability of the target domain samples on the source category, specifically by the following formula: ； wherein, denotes a softmax function, is a prediction classifier, ∈ {1,..., } is used to represent the source domain class index, represents the number of target domain samples; According to the source sample label y si Mapping category weights: ； where y si is the true label of the source sample x si . Integrate the sample weight into the adversarial loss function to realize feature alignment of the source domain public class features.

4. The adaptive fault diagnosis method of industrial equipment according to claim 3, characterized in that, The sample weight is integrated into the adversarial loss function, specifically: ； wherein L d (.) denotes the domain label prediction loss, and denote the real domain labels of the source domain and the target domain, respectively, denotes the discriminator, denotes the feature extractor, denotes the number of target domain samples, denotes the number of source domain samples, denotes the loss weight, denotes the i-th sample of the source domain, denotes the number of source domain samples.

5. The adaptive fault diagnosis method of industrial equipment according to claim 1, characterized in that, It also includes alleviating the excessive aggregation of unknown class features in the target domain by performing focal diversity processing on the target domain, and the focal diversity processing specifically includes L2 normalization on the target domain samples, and then calculating the cosine similarity of the sample pair; a focal weight mechanism is used to set a penalty weight for samples with high similarity.

6. The adaptive fault diagnosis method of industrial equipment according to claim 1, characterized in that, The general domain self-adaptive model also includes a classifier, which is used to predict the category label of a sample based on the feature vector generated by the feature extractor; for the source domain samples, the classifier is trained based on the labeled label to learn the mapping relationship between the features and the known categories; For the target domain samples, the classifier predicts the known categories to which the target domain samples belong based on the mapping relationship learned from the source domain.

7. An adaptive fault diagnosis system for industrial equipment, characterized by, It includes: A data acquisition module is configured to obtain equipment fault vibration data of an industrial scene; The equipment fault vibration data is divided into source domain samples and target domain samples; The classification processing module is configured to process the equipment fault vibration data based on a general field adaptive model to obtain a fault classification result of the general field, and perform fault diagnosis on the industrial equipment based on the fault classification result. In the general field adaptive model, the source domain is weighted, the weighted source domain samples and target domain samples are represented by a feature extractor, the feature distribution of the source domain and the target domain is distinguished by a discriminator based on the feature representation, the feature mean is taken by category to obtain a prototype vector of each known class, similarity calculation is performed on the target domain samples and the prototype vector of each known class, the sample with the lowest similarity to the prototype of the source domain is selected as an initial unknown class prototype according to a similarity threshold, the samples that are the nearest neighbors of the initial unknown class prototype in the target domain are assigned to the current unknown class based on a mutual nearest neighbor rule, all target domain samples are iteratively screened as new unknown class prototypes, the sample assignment process is repeated to obtain a preliminary clustering result of the target domain, and the preliminary clustering result is modified by calculating the sum of squared errors of the clustering result to output an adaptive classification result of the general field.

8. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-7. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when loaded by the processor, can execute the steps of the method of any one of claims 1 to 6.

Images