Double-path confrontation and progressive self-training driven cross-working-condition fault diagnosis method

Abstract

The invention discloses a double-channel confrontation and progressive self-training driven cross-working-condition open set domain adaptive fault diagnosis method. In order to suppress an interactive negative migration effect, a three-stage two-way adversarial progressive self-training network DAPN is designed, separation of unknown target samples and feature aggregation of known samples are realized through cooperation of a gradient inversion layer and a two-way adversarial discriminator, and inter-class discrimination and intra-class aggregation are obviously enhanced. When the performance of the DAPN is evaluated in two open set domain adaptive scenes, for open set proportions of different intensity domain offsets and changes, the method always keeps high classification accuracy and unknown class detection rate, and the performance is obviously superior to that of a comparison baseline. According to the method, two-way confrontation and a progressive self-training mechanism are combined, so that accurate classification of known faults and effective identification of unknown faults are realized.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a cross-condition fault diagnosis method driven by dual-path confrontation and progressive self-training. Background Technology

[0002] In recent years, the significant improvement in computing power and the accumulation of large-scale monitoring data have jointly driven the rapid development of data-driven deep learning (DL) methods {Zhu, 2022 #65;Rehman, 2025 #105}. Benefiting from its advantages such as independence from prior knowledge and powerful nonlinear feature extraction capabilities, deep learning has become an efficient solution in the field of rolling bearing fault diagnosis, demonstrating significant superiority in practical applications. Current mainstream methods include: Auto-encoders (AE): suitable for unsupervised feature reduction and reconstruction; Deep Belief Networks (DBN): extracting deep features through hierarchical pre-training; Convolutional Neural Networks (CNN): adept at processing local spatial features (such as vibration signal spectra); and Long Short-Term Memory Networks (LSTM): capturing long-term dependencies in time-series data. Generative Adversarial Networks (GANs) generate synthetic data to augment training samples. It's important to emphasize that the successful application of these methods relies on a key assumption: the source domain (training data) and the target domain (test data) must be independent and identically distributed. In other words, if the data distribution spaces of the training and test sets differ, the model's generalization ability will significantly decrease.

[0003] Fortunately, domain adaptation has made significant progress in mechanical fault diagnosis, especially demonstrating strong potential in addressing the domain offset problem. However, most existing methods are based on the closed-set assumption (i.e., the source and target domains have the exact same set of fault categories), which is often difficult to satisfy in real-world industrial scenarios because new and unknown fault types not included in the source domain may be encountered during deployment. Traditional Closed-Set Domain Adaptive (CSDA) methods often exhibit negative transfer effects when facing such unknown fault modes, leading to a significant deterioration in diagnostic performance. Fault diagnosis methods based on domain adaptation leverage knowledge from different domains to address these challenges. Domain Adaptive (DA), as an effective technique, has been used for cross-domain fault diagnosis. Early research mainly focused on edge distribution alignment, such as methods based on Maximum Mean Difference (MMD) and adversarial training methods, aiming to improve model generalization ability by minimizing the overall distribution difference between the source and target domains.

[0004] Subsequent Open-Set Domain Adaptive (OSDA) methods attempted to identify samples of unknown categories, but these methods typically failed to adequately address the domain offset problem simultaneously, limiting their applicability in cross-domain fault diagnosis scenarios. Some literature has addressed the OSDA problem and employed linear support vector machines to detect target outliers.

[0005] However, existing OSDA solutions still face two major challenges: (1) Challenges of Decision Boundary Uncertainty: Due to the lack of prior information on the openness of the target domain (the proportion of unknown class samples), it is difficult to set a universally applicable and robust decision threshold. Too tight a boundary (such as a high confidence threshold) can easily lead to a large number of unknown samples being missed in high openness scenarios; while too loose a boundary (such as a low threshold) may cause known samples to be misclassified as unknown in low openness scenarios. This sensitivity stems from the fact that traditional methods (such as OSBP) rely on a fixed threshold mechanism, while the dynamic changes in openness require the threshold to have adaptive capabilities in order to balance the trade-off between the missed detection rate and the false positive rate.

[0006] (2) Interactive Negative Transfer Challenge: In cross-domain open-set domain adaptation tasks, the known class accurate classification target and the unknown class detection target may interfere with each other due to gradient conflicts. When the domain offset is significant (e.g., MMD distance > 0.3), the distribution of known and unknown class samples in the feature space becomes mixed, causing the decision boundary to become blurred. Misclassified samples (e.g., samples from the shared space are misclassified into the private space) will destroy the stability of the feature distribution alignment, further distorting the learning direction of the decision boundary and forming a self-reinforcing negative feedback loop. This phenomenon is particularly prominent in adversarial training, where knowledge of the known classes in the source domain may be transferred to the unknown classes in the target domain through the incorrect guidance of the domain discriminator, thereby triggering interactive negative transfer. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a cross-condition fault diagnosis method driven by dual-path adversarial and progressive self-training. By using a gradient inversion layer and a dual-path adversarial discriminator to work together, the method can separate unknown target samples and aggregate features of known samples, thereby enhancing the inter-class discrimination and intra-class clustering. The classification accuracy and the detection rate of unknown classes are greatly improved, and the method can achieve accurate classification of known faults and effective identification of unknown faults.

[0008] To address the aforementioned technical problems, this invention provides a cross-condition fault diagnosis method driven by dual-path adversarial and progressive self-training, comprising the following steps: A dual-channel adversarial progressive self-training network is constructed, comprising a feature extraction module, a boundary construction module, and a feature alignment module. The feature extraction module is used to input the spectrum of the vibration signal into the feature generator G and extract shallow and deep features. The boundary construction module is used to remove unlabeled unknown samples from the target domain. The feature alignment module uses a dual-channel adversarial discriminator to align known samples in the source and target domains. A three-stage progressive self-training approach is adopted: First, the network parameters and weight coefficients of the loss function are initialized. A small batch of data is randomly sampled from the source and target domains for iterative training. The initial classification loss in the source domain is calculated. The feature space is initialized through an adversarial distribution alignment strategy to establish an accurate mapping of the source domain fault categories. Next, a boundary construction module is used to train the discriminator to detect and separate unknown fault types in the target domain. Finally, through adversarial training of the feature alignment module, only the known fault categories in the target domain are accurately aligned with the corresponding categories in the source domain, eliminating interference from unknown classes and forming a clear and distinguishable cross-domain feature distribution and an accurate decision boundary for unknown samples. Fault diagnosis is performed using a pre-trained dual-path adversarial progressive self-training network.

[0009] Furthermore, in the feature extraction module, a custom two-dimensional convolutional neural network is used as the feature extractor to automatically extract shallow and deep features from the source and target domains. The two-dimensional convolutional neural network consists of 5 convolutional layers and 1 fully connected layer. The input is a three-channel image, and the output is a 1024-dimensional feature vector.

[0010] Furthermore, the discriminator includes a coarse-grained discriminator. and fine particle size distinguisher The boundary construction module estimates the source risk term and open set risk term derived from the theoretical bound through cross-domain consistency scoring and uncertainty calibration; it uses confidence and entropy to calculate the classification loss for known data and the classification loss for unknown data.

[0011] Furthermore, we first use the K-Means clustering algorithm to generate pseudo-labels in the target domain, calculate the feature center points of the source domain and the target domain, and establish a cross-domain category mapping relationship through bidirectional similarity matching. We then calculate the average consistency score by using the classification accuracy of source domain samples at the center of the target domain and the classification accuracy of target domain samples at the center of the source domain to screen high-quality cross-domain matching pairs and avoid one-way matching bias. Secondly, through hierarchical adversarial training, the coarse-grained discriminator Used for aligning shallow features, learning globally invariant features, and fine-grained discriminators. This is used to achieve category-level semantic alignment by combining high-confidence target samples and uncertain samples for deep features; Finally, through entropy-weighted uncertainty sampling, samples are screened after predicting the entropy of the target domain and normalizing it. High-confidence samples are used for feature alignment, and "unknown class" constraints are imposed on low-confidence samples.

[0012] Furthermore, the feature alignment module introduces a dual-path adversarial discriminator for collaborative optimization. Dynamic weights are introduced into the dual-path adversarial loss, reducing the weights of low-confidence samples to suppress noise interference. This is based on the target discriminator. The output probability is used to construct the target sample. The weighting coefficients.

[0013] Furthermore, the adversarial distribution alignment strategy is as follows: By adopting a supervised learning paradigm and minimizing the cross-entropy loss function, the dual-path adversarial progressive self-training network establishes a mapping relationship from the original signal to the fault category.

[0014] Furthermore, the process of training the discriminator is as follows: During this stage, the source domain discriminator is trained. and target domain discriminator We construct decision boundaries from coarse-grained to fine-grained to distinguish between known and unknown fault categories in the target domain, which are used to jointly optimize the parameters of the feature generator G and the extended classifier C.

[0015] Furthermore, , , and These represent the feature generator G, the extended classifier C, and the target domain discriminator, respectively. Source domain discriminator The parameters, and the optimal parameter solution, are obtained through the following objective function: ; Represents classification loss, It is the source domain discrimination loss. It is the target domain discrimination loss. It is a loss based on entropy. This represents the weighting coefficient.

[0016] Furthermore, during adversarial training, an adversarial training strategy is adopted to optimize the parameters of the feature generator G and the source domain discriminator D, thereby aligning the feature distributions of the source domain and the target domain. This process only attracts known fault categories in the target domain to the corresponding categories in the source domain, thus eliminating interference from unknown fault categories.

[0017] The beneficial effects of this invention are: (1) This invention proposes a dual-path adversarial learning mechanism. By introducing two types of adversarial learning to cooperate with each other, the feature extractor is forced to learn height domain invariant features. These features retain discriminative ability, so that the model trained on the source domain can eventually adapt well to the target domain data.

[0018] (2) This invention designs a progressive self-training network and adopts a three-stage collaborative mechanism. Through the dynamic openness estimation module, the tightness of the decision boundary is dynamically adjusted to adapt to the domain offset of different openness. At the same time, a weighted adversarial distribution alignment module is introduced to drive the distribution alignment of the source domain and the target domain in the shared label space.

[0019] (3) This invention employs stepped channel compression combined with hierarchical regularization to suppress overfitting, uniformly discarding neurons and improving the generalization ability of samples. At the same time, a GRL layer is added after the designed five convolutional layers to reverse the gradient and work in conjunction with the designed dual-path adversarial discriminator to improve training efficiency and convergence stability. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the dynamic evolution of the DAPN of this invention; Figure 2 This is an implementation architecture diagram of the DAPN method of the present invention; Figure 3 This is a detailed parameter diagram of the network architecture of the present invention; Figure 4 This invention uses multiple methods to generate an average OS difference map for the PU bearing dataset. Figure 5 This invention uses multiple methods to generate an average OS difference map for a gearbox dataset. Figure 6 This is a diagram showing the classification results of existing methods on the OSDA task T5 of the PU bearing dataset; Figure 7 This is a diagram showing the classification results of the method of the present invention on the OSDA task T5 of the PU bearing dataset; Figure 8 This is a visualization of the OSBP method in cross-domain representation; Figure 9 This is a visualization of the FINN method in cross-domain representation; Figure 10 This is a visualization of the STA method in cross-domain representation; Figure 11 This is a visualization of the CMU method in cross-domain representation; Figure 12 This is the visualization result of the cross-domain representation method of the present invention; Figure 13 This is a similarity weight curve of the comparison model of this invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0022] Reference Figure 1As shown, an embodiment of the cross-condition fault diagnosis method driven by dual-path adversarial and progressive self-training of the present invention aims to achieve accurate classification of known faults and effective identification of unknown faults. Based on minimizing the theoretically derived upper bound of the error, this method synergistically optimizes the open-domain detection loss and the domain-invariant learning loss to achieve separability between class levels and tightness within class levels. To further suppress the interactive negative transfer effect, a three-stage progressive self-training network is designed: In the initial stage, an adversarial discriminator is used to calculate the domain classification probabilities of shallow features in the source and target domains, and the accuracy of source domain classification is measured by calculating the cross-entropy loss. Secondly, in the second stage, a dynamic decision boundary is adaptively generated to accurately distinguish unknown samples in the target domain, and the separability of outlier data (between classes) is enhanced by iteratively optimizing the source domain risk and open set risk. In the third stage, a dual-path adversarial mechanism is introduced to give the model the ability to learn discriminative representations from imbalanced data, learn domain-invariant features sharing the label space across the source and target domains, improve the aggregation of intra-class samples, and ensure robust generalization performance under different openness levels.

[0023] The dynamic evolution process of the proposed Dual-Adversarial Progressive Self-Training Network (DAPN) is as follows: Figure 1 As shown: Figure 1 (a) Initial state: In the open set scenario, the source and target domains share known fault categories (such as triangles and pentagons), while the target domain contains additional unknown fault categories (heart-shaped and sun-shaped). An adversarial distribution alignment strategy is used to initialize the feature space. Figure 1 (b) Second stage: Establish the classification boundary between known fault categories and unknown fault categories to realize basic diagnostic functions; secondly, construct two-level decision boundaries with coarse and fine granularity to separate known and unknown class samples in the target domain. Figure 1 (c) Third stage: By adversarial distribution alignment, the known class samples of the source domain and the target domain are brought closer to the shared feature space, thereby enhancing intra-class compactness. Figure 1 (d) Convergence state: Through iterative optimization by the unknown separation module and the attraction optimization module, the unknown class samples in the target domain are pushed away from the shared space, while the known class samples are closely aggregated with the same class samples in the source domain, forming a clear and identifiable cross-domain feature distribution.

[0024] In the problem of defining open sets, it is generally assumed that there exists source domain data. With existing target domain data Its distribution characteristics satisfy the requirement that the source domain labeled data follows a joint distribution. ,in For the sample size, For the source domain feature space, This represents the category label space. The unlabeled data in the target domain follows a marginal distribution, i.e. Open set classification tasks require training a target domain classifier. , It needs to accurately classify known classes, and classify samples in the target domain that belong to known classes. Output the correct label The ability to detect unknown samples must meet the requirement of detecting unknown samples in the target domain. The output is uniformly labeled "Unknown fault type". In the label space definition, the source domain label space... It consists of c known fault categories. The target domain label space... New Indicates an unknown class.

[0025] This invention aims to construct a deep diagnostic model that accurately classifies test samples from known categories to... If the category is known, there is no need to classify the unknown fault category; simply reject test samples of the unknown category into the "unknown fault type" category.

[0026] For the theoretical error upper bound, given the hypothesis space Satisfying the mild conditions (i.e., classifier) (The Lipschitz constant is a bounded value) Expected risk in the target domain The upper bound can be expressed as: ; Here, k is a constant called the shared error. Indicates the number of known categories. This indicates the proportion of unknown class samples in the target domain. The shared error coefficient reflects the minimum inherent deviation after the source and target domains are aligned. By performing a four-dimensional decomposition of the upper bound of the error, we can see that the performance of classifier C in the target domain is constrained by four joint constraints.

[0027] Source risk items This represents the classification loss for known samples in the source domain. Since the source domain does not contain unknown samples, minimizing the source domain classification error can be directly optimized by optimizing this term.

[0028] Difference distance : Distribution of known target domains in the quantization feature space Global distribution of the target domain The difference. Reducing this distance requires learning domain-invariant features.

[0029] Shared error When conditional distribution When inter-domain offset exists (even if the edge distribution is aligned). Significantly increased.

[0030] Open set risk correction item : Describes the inter-domain differences in classification performance for unknown classes, where This represents the partial risk of the unknown class in the model.

[0031] For domain-invariant features, the optimization objective is to minimize the feature distribution difference using a generative adversarial network: ; Where G is the feature generator, which extracts domain-invariant features. D is the domain discriminator, which drives feature space alignment. Domain-specific features are eliminated by inducing domain confusion.

[0032] Risks related to open sets , is defined as: ; It has a two-way adjustment mechanism, positive item Promote the identification of unknown samples in the target domain, negative terms Constrain known samples in the source domain to prevent them from being misclassified as unknown classes. Subject to target domain openness (Unknown category proportion) and unknown category partial risk Joint influence.

[0033] For the dual-path adversarial discriminant architecture, the feature generator G and the domain discriminator D1 learn domain-invariant features through a minimax game, focusing on the overall distribution alignment between the source and target domains. The optimization objective is: ; The class-aware discriminator D2 calculates the class-conditional similarity between the target sample and the source domain. A higher output value indicates a greater similarity to the known class distribution of the source domain. The class-aware confidence score output by D2 provides feedback to the boundary construction module, dynamically sharpening the decision boundary for known / unknown classes. It is defined as: ; Based on the above, the core architecture of the Dual-Path Adversarial Progressive Self-Training Network (DAPN) is described in detail below. Figure 2 As shown: The core architecture consists of three parts: a feature extraction module, a boundary construction module, and a feature alignment module. First, in the feature extraction module, the spectrum of the vibration signal is input into the feature generator G to extract shallow and deep features. Second, in the boundary construction module, the source risk term and open set risk term derived from the theoretical bound are estimated through cross-domain consistency scoring and uncertainty calibration; confidence and entropy are used to calculate the classification loss for known data and the classification loss for unknown data. Third, in the feature alignment module, a dual-path adversarial discriminator is proposed. The feature generator G and the domain discriminator D1 learn transferable features through minimax adversarial training; simultaneously, the class-aware discriminator D2, which does not participate in adversarial training, is responsible for evaluating the similarity between a single target sample and the source domain. This avoids gradient conflicts caused by a single discriminator simultaneously optimizing alignment and separation. The confidence level is calculated by combining the dual-path adversarial loss, and known / unknown samples are divided based on the confidence level. For unknown samples, fewer weights are assigned, and additional constraints are imposed on low-confidence samples to prevent negative transfer.

[0034] Specifically, refer to Figure 3 As shown, in the feature extraction module described above, a custom two-dimensional convolutional neural network is used as the feature extractor to automatically extract shallow and deep features from the source and target domains. This network is a feature extractor consisting of 5 convolutional layers and 1 fully connected layer. The input is a three-channel image, and the output is a 1024-dimensional feature vector. Its core design idea is to progressively abstract image features through multiple convolutions, ultimately compressing them into a high-dimensional semantic representation. The 5 convolutional layers (conv1 to conv5): progressively extract spatial features, with the number of channels decreasing (128→64→32→32→16). Conv1: 128 3x3 convolutional kernels, increasing the number of output channels from 3 to 128, extracting low-level edge features. Conv2~Conv4: progressively reduce dimensionality (128→64→32→32 channels). Conv5: 16 3x3 convolutional kernels, reducing the number of output channels to 16, generating high-dimensional semantic features. ReLU is used after each convolutional layer to introduce non-linearity, enhancing feature expressiveness and solving the gradient vanishing problem. Dropout is added after each layer to prevent overfitting. The first four layers have a p=0.25 value, and the last layer has a p=0.8 value. The last layer uses a high dropout rate to extract high-level semantic features, enhancing generalization ability. Then, a fully connected layer (fc) compresses the spatial features into a 1024-dimensional vector. Finally, a gradient inversion layer (GRL) is introduced to encourage the feature extractor to generate domain-invariant features and work in conjunction with the designed dual-path adversarial discriminator during adversarial training, achieving end-to-end optimization.

[0035] In the aforementioned boundary construction module, unlabeled unknown samples are pushed out of the target domain. First, K-Means clustering is used to generate pseudo-labels in the target domain, and feature centroids of the source and target domains are calculated. Cross-domain category mapping is established through bidirectional similarity matching (e.g., cosine similarity matrix). The average consistency score is calculated using the classification accuracy of source domain samples at the target domain center and the classification accuracy of target domain samples at the source domain center, filtering high-quality cross-domain matching pairs and avoiding one-way matching bias. Secondly, to more accurately calculate openness, this paper employs hierarchical adversarial training and a coarse-grained discriminator. Used to align shallow features and learn globally invariant features; fine-grained discriminator This method is used to target deep features, combining high-confidence target samples and uncertain samples to achieve category-level semantic alignment. Finally, entropy-weighted uncertainty sampling is used, and samples are selected after normalization based on the target domain predicted entropy (p1_entropy). High-confidence samples are used for feature alignment, while low-confidence samples are subject to an "unknown class" constraint to prevent misalignment. Higher entropy indicates higher uncertainty in class distribution, thus calibrating similarity. Represented as: ; in Represents the probability of a known class. This represents the features extracted from the source domain. For the features in the target domain, the binary discriminator outputs a probability vector. Can be regarded as source category The prediction confidence level is determined. This method is effective for target data with known label spaces (due to their shared source domain structure), but for target data with unknown label spaces, their unique structural features are not covered by the source domain, resulting in weak discrimination ability and biased similarity estimation. A self-supervised uncertainty calibration technique based on entropy analysis is introduced. Entropy value. To measure the smoothness of the probability distribution (higher entropy indicates greater uncertainty), the calibrated similarity is defined by combining normalized entropy with a confidence term: ; By jointly optimizing a multi-objective loss function, classification loss, adversarial loss, and uncertainty constraints, the generalization performance of the model in the target domain is improved. The loss function is defined as: ; in, It is classified loss, / Is it coarse / fine granularity to combat loss? It is an uncertain sample loss. It is the weighting coefficient.

[0036] The aforementioned feature alignment module attracts known fault type samples from the target domain, thereby promoting alignment between known samples in the source and target domains and avoiding negative transfer caused by unexpected alignment between known samples in the source domain and unknown samples in the target domain. This module introduces a dual-path adversarial discriminator for collaborative optimization, incorporating a dynamic weight A in the dual-path adversarial loss. The weight of low-confidence samples is reduced to suppress noise interference. Based on the target discriminator... The output probability is used to construct the target sample. Weighting coefficients: ; This indicates that the samples come from a shared label space (known class) and should be the focus of learning in distribution alignment. This indicates that the sample belongs to a private label space (unknown class), and its alignment interference needs to be suppressed. The total loss function integrates three parts: ; in, Indicates classification loss; This indicates the losses from dual-path confrontation; Indicates the loss of an uncertain sample. It is the weighting coefficient.

[0037] After the dual-path adversarial progressive self-training network DAPN is established, network training is still required. The DAPN network adopts a three-stage progressive self-training process as follows: First, initialize all network parameters and the weights of the loss function. A mini-batch of data is randomly drawn from both the source and target domains for training in this iteration. The initial classification loss in the source domain is calculated. This ensures the model can correctly classify features from the source domain. The feature space is initialized using an adversarial distribution alignment strategy to establish an accurate mapping between source domain fault categories. Next, a dedicated discriminator is trained in the boundary construction module to effectively detect and separate unknown fault types in the target domain. Finally, adversarial training in the feature alignment module precisely aligns only known fault categories in the target domain with their corresponding categories in the source domain, eliminating interference from unknown classes. This structured training process significantly optimizes the theoretical error boundary of open-set domain adaptation, effectively suppresses negative transfer effects, and ultimately forms a clearly discernible cross-domain feature distribution and accurate decision boundaries for unknown samples.

[0038] Specifically, first perform adversarial distribution alignment: In the first stage of model training, the focus is on jointly optimizing the core components of the feature generator G and the extended classifier C. The optimization objective is to maximize classification accuracy for the health state of the source domain machine (i.e., predefined fault categories). This stage employs a supervised learning paradigm, minimizing the cross-entropy loss function to establish a precise mapping relationship between the original signal and the fault category, laying the foundation for feature discrimination in subsequent stages.

[0039] Then the boundary construction phase begins: In this stage, a dedicated source domain discriminator is trained. and target domain discriminator This involves constructing decision boundaries from coarse-grained to fine-grained to distinguish between known and unknown fault categories in the target domain. Let... , , and These represent the feature generator G, the extended classifier C, and the target domain discriminator, respectively. Source domain discriminator The optimal parameter solution can be obtained by the following objective function: Represents classification loss, It is the source domain discrimination loss. It is the target domain discrimination loss. It is a loss based on entropy. Indicates the weighting coefficient; Finally, there is the feature alignment stage: In this stage, an adversarial training strategy is employed to optimize the parameters of the feature generator G and the domain discriminator D, achieving feature distribution alignment between the source and target domains. This process specifically emphasizes attracting only known fault categories from the target domain to their corresponding categories in the source domain (i.e., excluding interference from unknown fault categories). Let... , ,and represent the parameters of the feature generator G, the extended classifier C, and the bi-channel adversarial discriminator, respectively.

[0040] ; ; The DAPN method proposed in this application significantly optimizes the theoretical error upper bound of the Open Set Domain Adaptive (OSDA) problem through a three-stage progressive self-training mechanism, effectively suppressing the negative transfer effect caused by unknown fault categories. Based on the feature space initialization using an adversarial distribution alignment strategy, through continuous iterative optimization of boundary construction and feature alignment, unknown class samples in the target domain are pushed out of the shared space, while known class samples are tightly aggregated with similar samples in the source domain. Therefore, the decision boundary for unknown data can be constructed more accurately.

[0041] The method of this application will be described in detail below using two datasets.

[0042] The first dataset comes from the University of Paderborn's Rolling Bearing Dataset, which contains bearing samples with artificially created faults and actual operational damage. This dataset uses piezoelectric accelerometers to collect bearing housing vibration signals at a sampling frequency of 64 kHz. Multi-condition data acquisition was achieved by adjusting the drive system speed, testing the bearing radial force, and measuring the transmission load torque. Eight bearings were selected for accelerated life testing to obtain real-world damage data. The operating conditions for this dataset are shown below: Rotational speed: 900 rpm / 1500 rpm, load torque: 0.7 Nm / 0.1 Nm, radial force: 1000 N / 400 N. O: outer ring fault; I: inner ring fault; IO: inner and outer ring faults; H: healthy. S: single fault; R: recurring fault; F: fatigue pitting; P: plastic deformation.

[0043] The second dataset, derived from the PHM 2009 Data Challenge, consists of a planetary gearbox dataset and includes three drive shafts, four gears, and six bearings. The experiment targeted different fault types in two sets of gears (spur gears and helical gears). The acquired signals included two channels of vibration signals (acquired by an accelerometer) and one channel of speed signals (synchronously recorded by a speed sensor). Signals under all healthy conditions were continuously acquired for 4 seconds at a sampling frequency of 66.67 kHz. The dataset covers six healthy states of the planetary gearbox; detailed fault descriptions are shown in the table below. Shaft speed: 30 / 35 / 40 / 45 / 50 Hz, Load: Low / High. Source domain condition: Load torque: 0.1 Nm, Radial force: 1000 N. Target domain condition: Load torque: 0.7 Nm, Radial force: 400 N.

[0044] Hyperparameter settings and experimental framework: The hyperparameters of all comparison methods were determined based on previous research and experimental requirements to ensure satisfactory performance. PyTorch was used as the model training framework, and an NVIDIA 4070Ti GPU was utilized to accelerate the computation process.

[0045] Comparison Method Design: This method introduces the concept of different technical approaches and integrates six cutting-edge Open Set Domain Adaptation (OSDA) methods to verify the effectiveness of the proposed method in fault diagnosis tasks. To ensure fairness in the comparison, all methods are implemented using similar network structures.

[0046] (1) The core mechanism of OSBP (Open Set Adaptive Method Based on Backpropagation) is to construct a decision boundary that can identify unknown class target samples by jointly optimizing the classifier and the generator. Specifically, the classifier training objective is to establish an explicit decision boundary between the source domain samples and the target domain samples; while the generator training objective is to drive the target domain samples away from this boundary, thereby enhancing the model's ability to distinguish unknown class samples.

[0047] (2) DATLN (Deep Adversarial Transfer Learning Network) innovatively proposes a solution to the open set problem of diagnosing knowledge transfer in industrial scenarios. This method designs an adversarial classifier that aligns known class samples in the source and target domains within a shared feature space, while simultaneously detecting unknown class samples.

[0048] (3) IW-OSDA (Instance-weighted Open Set Domain Adaptation) significantly improves the detection capability of unknown fault samples in the Open Set Domain Adaptation (OSDA) task by integrating an anomaly classifier into an adversarial network framework. This method innovatively utilizes domain dissimilarity metrics to generate instance-level weights to quantify the similarity between target domain samples and known categories in the source domain.

[0049] (4) CMU (Calibrated Multiple Uncertainty Method) proposes a novel transferability metric for detecting out-of-distribution data. This metric achieves a comprehensive evaluation of outlier samples by fusing three complementary uncertainty quantification metrics, based on entropy, confidence, and consistency defined by the conditional probability calibrated by the multi-classifier ensemble model.

[0050] (5) FTNN (Feature-Based Transfer Neural Network) extracts transferable features through multi-layer domain adaptation and pseudo-label learning. This method minimizes the maximum mean difference (MMD) between different layers of the convolutional neural network (CNN) to align the distributions of the source and target domains.

[0051] (6) STA (Separate to Adapt) uses a two-stage progressive network architecture to solve the open set transfer learning problem: the first stage (separation stage) identifies and separates unknown category samples in the target domain by mining the features of the source domain data; the second stage (adaptation stage) realizes the distribution adaptation between the source domain and the target domain within the known category range.

[0052] (7) DAPN (the proposed method). The Adam optimizer is used, combined with weight decay to prevent overfitting. The initial learning rate is uniformly set to 5e-4. In the early stage, a higher learning rate is used to achieve rapid convergence. In the later stage, the learning rate is reduced to fine-tune the parameters and avoid oscillation. That is, the learning rate decays significantly with the increase of training steps.

[0053] In addition, all training batches are fixed at 40, and dynamic openness is used for calculation.

[0054] Ten open-set domain adaptation tasks were constructed based on bearing and gearbox datasets. These include five bearing OSDA tasks and five gearbox OSDA tasks. Bearing OSDA task design: The PU bearing dataset was selected. By progressively increasing the openness gap (i.e., the difference in the proportion of unknown categories) between the source and target domains, the diagnostic robustness and accuracy of different methods in scenarios with increasing openness were systematically evaluated. PU bearing dataset open-set domain adaptation tasks: The load torque is 0.1 N·m, and the radial force is 1000 N. The operating parameters of the target domain are: load torque 0.7 N·m, and radial force 400 N.

[0055] Gearbox OSDA Task Design: Five tasks are constructed based on the Gearbox dataset. By using an increasing inter-domain openness gap, the adaptability of models to unknown fault categories is quantitatively compared. The source and target domain samples exhibit a significant imbalance in their health state distributions, aiming to test the model's generalization performance under data distribution bias. PHM09 Gearbox Dataset Open-Set Domain Adaptation Task: The load torque is low, and the shaft speed is 40Hz; the target domain operating parameters are: high load torque and shaft speed of 45Hz.

[0056] In the field of OSDA research, the design of performance evaluation metrics is crucial for accurately measuring algorithm performance. Based on existing research, two core evaluation metrics are used to quantify the performance of open-set domain adaptive tasks. The definitions of each metric are as follows: The number of known class samples correctly classified in the target domain; this metric reflects the model's ability to identify samples of known classes. The number of unknown class samples in the target domain that are correctly classified (i.e. predicted as "Unknown"). This metric measures the model's ability to reject samples of unknown class. The total number of known class samples in the target domain; The total number of unknown class samples in the target domain; Known class accuracy is a metric that specifically evaluates the model's classification accuracy for samples of known classes in the target domain. It reflects the model's ability to preserve known knowledge; a higher accuracy indicates better performance. The value indicates that the model can effectively retain the knowledge learned in the source domain and apply this knowledge effectively in the target domain. The formulas used are shown below: ; Overall accuracy in the target domain: The target domain typically includes new categories not seen in the source domain, requiring the model to have a good balance between identifying unknown classes and preserving known classes. This comprehensive metric considers both the model's ability to classify known classes and its ability to identify unknown classes, fully reflecting the model's overall performance in an open-set environment. The formula used is shown below: ; These two metrics together constitute a complete evaluation framework for open-set domain adaptive tasks: Focusing on the model's knowledge retention capabilities, and This measures the overall adaptability of the model. In practical applications, both metrics need to be considered simultaneously to ensure that the model is neither overly conservative, resulting in a low recognition rate for unknown classes, nor overly aggressive, resulting in a decrease in accuracy for known classes.

[0057] Analysis of diagnostic accuracy for the PU bearing dataset: The diagnostic results of the PU bearing dataset are shown in the table below: The above includes the overall accuracy within the target domain. Accuracy with only known class To more vividly illustrate the average of each method and the proposed method... Differences, average of PU bearing dataset Differences such as Figure 4 As shown; to reduce the impact of randomness, the average accuracy was calculated through ten independent trials as a performance evaluation index.

[0058] The proposed DAPN method demonstrates general superiority across all relevant tasks. It should be noted that when the evaluation focuses only on the accuracy of known classes (…),… When considering all categories (including unknown categories), some methods may outperform DAPN; however, when considering the overall performance of all categories (including unknown categories), the performance of these methods drops significantly. For example, in bearing task T5, if unknown categories are ignored, the FINN model achieves an accuracy of 100%, outperforming DAPN. However, when new unknown fault data from the target domain are included, the model's generalization ability drops sharply due to misclassification of unknown faults, and its overall accuracy drops to 36.6%. OSBP and DATNN models also have the same problem. When there is no interference from unknown class samples, the recognition rate of known class samples is good, but when there is interference from unknown samples, the model's recognition rate drops significantly. This indicates that unknown fault sample categories can bring negative transfer to the model, affecting its performance.

[0059] Figure 6 and Figure 7 As shown, the classification performance of various models under different degrees of openness on bearing task T5 is further compared. In complex mechanical systems, equipment may frequently experience new unknown faults (such as unrecorded sudden faults). High openness simulates such scenarios, testing the model's diagnostic ability when "known categories are sparse and unknown categories dominate". This method achieves optimal performance on an industrial dataset with an openness as high as 0.7, significantly improving the diagnostic accuracy in high-difficulty scenarios. The results show that traditional transfer learning models are all affected by the negative transfer effect caused by unknown fault categories. The main reason for the limited performance of such standard models is that they adopt a global domain alignment strategy but do not consider the interference of unknown samples in the target domain. This non-discriminatory alignment strategy leads to the misclassification of unknown fault samples as known categories, and forced alignment will confuse the boundaries between known and unknown classes, reducing the model's generalization ability.

[0060] The aforementioned issues severely limit the robustness of FINN and DATLN in Open Set Domain Adaptation (OSDA) scenarios, and their performance degradation becomes more significant as the openness of the target domain increases. Experimental results show that this proposed method significantly improves clustering accuracy compared to traditional methods. Furthermore, the standard deviation of this method's performance is generally low across different tasks, indicating good convergence in multiple experimental runs.

[0061] Diagnostic accuracy analysis of the PHM gearbox dataset: The diagnostic results of the PHM09 gearbox dataset are as follows: The above includes the overall accuracy within the target domain. Accuracy with only known class The average value of the PHM09 gearbox dataset Differences such as Figure 5As shown in the figure, several findings are consistent with observations on the bearing dataset. DAPN achieved the highest diagnostic accuracy in most tasks, and in the gearbox task T5, the FINN model achieved 100% accuracy, outperforming DAPN, if unknown categories were ignored. However, when new unknown fault data from the target domain were included, the model's generalization ability dropped sharply due to misclassification of unknown faults, and its overall accuracy decreased to 34.1%.

[0062] The high-dimensional feature representations learned by different methods are reduced in dimensionality and visualized using the t-SNE technique (see related results). Figure 8-12 Analysis shows that in the visualization results of cross-domain representations, both OSBP and FINN methods exhibit significant overlap between known and unknown class samples in the target domain. This mixing phenomenon indicates that the model struggles to effectively distinguish different failure modes, posing a challenge to the classifier's discrimination accuracy and leading to a significant performance degradation in the Open Set Adaptive Task (OSDA). The baseline methods STA and CMU employ different weights to distinguish unknown failure samples, which is related to the approach in this application. Ideally, samples sharing the health state with the source domain should receive larger weights, while the weights of target outlier samples should approach zero. However, feature visualization results show reduced discriminability of the baseline methods, specifically due to the following limitations of STA. This method uses a two-stage discriminator structure, only able to identify outlier samples with significant differences, treating them as homogeneous categories. This strategy fails to fully utilize the intrinsic structural features of outlier samples (such as local similarity), leading to the neglect of anomalous samples with insignificant divergent features. This results in some outlier samples with similar features being identified as known failure categories, leading to an overall performance degradation. The CMU method constructs target instance weights by fusing statistical measures such as confidence, entropy, and consistency, aiming to distinguish normal samples from outliers. However, when outliers are highly similar to known classes in feature distribution or structure, the discriminative power of this mechanism significantly decreases. In open environments, structural similarity between outliers and known classes is unavoidable, and the CMU method relies excessively on surface statistical features of probability distributions, lacking deep modeling of semantic structure and cross-modal heterogeneity. The results of our proposed method clearly show that unknown class samples in the target domain are well isolated from samples in the source domain, while known class samples in the target domain are effectively aggregated and aligned with their corresponding known class samples in the source domain. These visualizations strongly demonstrate that the method proposed in this application can effectively extract domain-invariant feature representations of known classes and successfully filter out interference from unknown class samples during domain adaptation.

[0063] For ablation experiments, to evaluate the effectiveness of the dual-path adversarial discriminator—specifically, the domain discriminator D1 based on calibrated similarity, the adaptive class-aware discriminator D2 based on domain consensus scoring, and the alignment module based on weighted adversarial distribution—three different models based on the DAPN method were designed for comparative experiments. The specific models included the domain discriminator D1 without calibrated similarity, the adaptive class-aware discriminator D2 without adaptive similarity, and the alignment module without weighted adversarial distribution. The changes in similarity weights learned by the four models were investigated on the open-set domain adaptive task T3 of the PU bearing dataset and the open-set domain adaptive task T3 of the PHM09 gearbox dataset to evaluate the effectiveness of DAPN in constructing an accurate hyperplane separating known and unknown classes. To obtain the changes in similarity weights throughout the training process, the similarity weights for each class were averaged over one iteration. Cosine similarity was used to calculate the similarity weight norm during the training process; the horizontal axis represents the number of training iterations, and the vertical axis represents the L2 norm (normalized value) of the weights. Figure 13 As shown in the figure. In the open-set domain adaptive task T3 of the PU bearing dataset, the step size was set to 40. In the comparison with the comparative methods, it can be seen that the weight of Class 0 (outer ring single-point fault) reached the full score of 1.0, giving the highest priority to key faults and effectively suppressing unknown classes (weight 0.9). It achieved full stability in 15 steps and quickly reached convergence, making the real-time diagnostic system respond faster. In the open-set domain adaptive task T3 of the PHM09 gearbox dataset, the step size was set to 100. The edge computing advantage of DAPN was utilized, and stability was achieved in 15 steps, saving 60% of training energy compared with (d), (f), and (h) in the figure. Maintaining the weight of key faults at 0.95, there is no risk of overfitting. In (f), the model's Class 0 continues to decrease, and the diagnostic ability deteriorates. In (h), the model's Class 0 frequently jumps, indicating a risk of real-time diagnosis. This shows that the model can more accurately separate unknown classes from the target domain. In summary, the comparative experiments above demonstrate that DAPN exhibits significant advantages in three aspects: the proposed domain discriminator D1 based on calibration similarity, the adaptive class-aware discriminator D2 based on domain consensus scoring, and the alignment module based on weighted adversarial distribution. The model integrating all modules achieves optimal performance, while removing any module leads to a performance degradation across all tasks.

[0064] The present invention also analyzes the computational efficiency: The table below lists the average training and testing times of different comparison methods on the open-set domain adaptive task T5 of the PU bearing dataset and the open-set domain adaptive task T1 of the PHM09 gearbox dataset. This study compares the average training and testing times of different methods on the PU bearing and PHM09 gearbox datasets (note: testing time is the sum of prediction times for all test samples). Experimental results show that the DAPN method maintains high training speed (especially the fastest on the PHM09 dataset) while also exhibiting good efficiency in the testing phase (fastest on the PHM09 dataset, second only to OSBP on the PU dataset). The proposed method shows a significant advantage in training efficiency, not only with shorter execution time but also with stronger ability to identify unknown classes.

[0065] Computation time (s) of each comparison method and the proposed method: This invention proposes a cross-condition open-set domain adaptive fault diagnosis method driven by dual-path adversarial and progressive self-training. To suppress the interactive negative transfer effect, a three-stage dual-path adversarial progressive self-training network (DAPN) is designed. Through the coordinated use of a gradient inversion layer (GRL) and a dual-path adversarial discriminator, the method achieves separation of unknown target samples and feature aggregation of known samples, significantly enhancing inter-class discrimination and intra-class clustering. Finally, the performance of DAPN is evaluated in two open-set domain adaptive scenarios. Comparative experimental results show that this method maintains good classification accuracy and unknown class detection rate even with changes in domain offset intensity and open set ratio. Compared with suboptimal methods, DAPN shows significant performance improvement on most tasks. In the open-set domain adaptive task T3 of the PU bearing dataset, the average OS is 3.1% higher than the second-place CMU. In the open-set domain adaptive task T5 of the PHM09 gearbox dataset, the average OS is 5% higher than the second-place IW-OSDA, significantly outperforming the baseline. DAPN maintains leading performance across all tasks on two different datasets, demonstrating strong generalization ability and stability. In contrast, other methods such as CMU and IW-OSDA perform well on some tasks but lack consistency. The proposed method achieves accurate classification of known faults and effective identification of unknown faults by combining dual-path adversarial and progressive self-training mechanisms.

Claims

1. A cross-condition fault diagnosis method driven by dual-path adversarial and progressive self-training, characterized in that, The method comprises the following steps: First, a two-way progressive self-training network is constructed: The feature extraction module, the boundary construction module and the feature alignment module are included, the feature extraction module is used for inputting the frequency spectrum of the vibration signal into the feature generator G and extracting shallow features and deep features; The boundary construction module is used for removing unmarked unknown samples from the target domain; the feature alignment module uses a two-way adversarial discriminator to align known samples in the source domain and the target domain; Then, three-stage progressive self-training is adopted: first, the network parameters and the weight coefficients of the loss function are initialized, a small batch of data is randomly extracted from the source domain and the target domain for iterative training, the initial classification loss on the source domain is calculated, the feature space is initialized through the adversarial distribution alignment strategy, and the accurate mapping of the source domain fault categories is established; then, the discriminator is trained by using the boundary construction module, which is used for detecting and separating unknown fault types in the target domain; finally, through the adversarial training of the feature alignment module, only the known fault categories in the target domain are accurately aligned with the corresponding categories in the source domain, the unknown class interference is excluded, and clear and distinguishable cross-domain feature distribution and accurate unknown sample decision boundary are formed; Finally, the trained two-way progressive self-training network is used for fault diagnosis.

2. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 1, wherein, In the feature extraction module, a self-defined two-dimensional convolutional neural network is used as a feature extractor for automatically extracting shallow features and deep features from the source domain and the target domain, the two-dimensional convolutional neural network is composed of 5 convolutional layers and 1 fully connected layer, the input is a three-channel image, and the output is a 1024-dimensional feature vector.

3. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 1, wherein, The discriminator includes a coarse-grained discriminator and a fine-grained discriminator ; a boundary construction module estimates a source risk term and an open set risk term under a theoretical bound derivation through cross-domain consistency scoring and uncertainty calibration; confidence and entropy are used to calculate classification loss of known data and classification loss of unknown data.

4. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 3, wherein, Firstly, K-Means clustering algorithm is used to generate pseudo labels in the target domain, the feature center points of the source domain and the target domain are calculated, and the cross-domain category mapping relationship is established through bidirectional similarity matching; the average consistency score is calculated through the classification accuracy of the source domain samples in the target domain center and the classification accuracy of the target domain samples in the source domain center, and high-quality cross-domain matching pairs are screened to avoid one-way matching deviation; Second, through hierarchical adversarial training, coarse-grained discriminator For aligning shallow features, learning global domain invariant features, fine-grained discriminator For deep features, combining high-confidence target samples and uncertain samples to achieve class-level semantic alignment; Finally, through entropy weighted uncertainty sampling, samples are screened based on the target domain prediction entropy after normalization, high confidence samples are used for feature alignment, and low confidence samples are subjected to "unknown class" constraint.

5. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 4, wherein, The feature alignment module introduces a two-way adversarial discriminator to cooperatively optimize, introduces a dynamic weight in the two-way adversarial loss, reduces the weight of low confidence samples, and suppresses noise interference. Based on the output probability of the target discriminator , the weight coefficient of the target sample is constructed.

6. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 1, wherein, The adversarial distribution alignment strategy is as follows: In the supervised learning paradigm, the two-way progressive self-training network is used to minimize the cross-entropy loss function to establish the mapping relationship from the original signal to the fault category.

7. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 6, wherein, The process of training the discriminator is as follows: In this stage, the source domain discriminator and the target domain discriminator are trained to build the decision boundary from coarse to fine to distinguish the known and unknown fault classes in the target domain for jointly optimizing the parameters of the feature generator G and the extended classifier C.

8. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 7, wherein, Let , , and denote the parameters of the feature generator G, the expansion classifier C, the target domain discriminator and the source domain discriminator , respectively, and the optimal parameter solution is obtained by the following objective function: ； represents a classification loss, is a source domain discriminative loss, is a target domain discriminative loss, is an entropy-based loss, represents a weight coefficient.

9. The dual path adversarial and progressive self-training driven cross- operating condition fault diagnosis method of claim 1, wherein, In the adversarial training process, the parameters of the feature generator G and the source domain discriminator D are optimized by using the adversarial training strategy to realize the alignment of the feature distribution of the source domain and the target domain, and only the known fault categories in the target domain are attracted to the corresponding categories in the source domain, that is, the interference of unknown fault categories is excluded.

Images