Aug 2026· Measurement science and technology· 0 citations
TL;DR
A novel and robust Source-Free Domain Adaptation (SFDA) framework equipped with post-hoc explainability for bearing fault diagnosis is proposed, and a post-hoc visualization mechanism is incorporated to explain the model's decision-making process, enhancing the transparency and credibility of the diagnosis for practical industrial deployment.
Abstract
While conventional fault diagnosis excels in steady-state scenarios, reliable detection under variable operating conditions remains a critical challenge in modern industrial drive systems. Although domain adaptation techniques address cross-domain distribution shifts, their deployment is often hindered by data privacy concerns and the impracticality of labeling massive target data in real-world factories. Furthermore, standard methods typically require concurrent access to source and target data, which violates the data isolation protocols common in industrial scenarios. To overcome these barriers, this paper proposes a novel and robust Source-Free Domain Adaptation (SFDA) framework equipped with post-hoc explainability for bearing fault diagnosis. First, to combat complex environmental interference and uneven sample distributions, we introduce a Gaussian noise-based stochastic perturbation strategy, forcing the model to learn robust decision boundaries. Second, high-confidence pseudo-labels are mined as reliable supervision signals. To fundamentally optimize the feature manifold, we introduce a contrastive learning-based representation strategy. Specifically, a mixed loss combining Supervised Contrastive loss and dot product loss is proposed to explicitly enhance feature discriminability while preventing feature collapse, achieving a delicate balance between intra-class compactness and overall diversity. Finally, a post-hoc visualization mechanism is incorporated to explain the model's decision-making process, thereby enhancing the transparency and credibility of the diagnosis for practical industrial deployment. Extensive experiments on two publicly available bearing fault datasets demonstrate that the proposed method significantly outperforms existing SFDA-based baselines. On the PU dataset, our approach achieves an average diagnostic accuracy of 99.11%, yielding a performance gain of 1.86% over the state-of-the-art SDALR method. Similarly, on the JNU dataset, the proposed framework attains an average accuracy of 98.36%, demonstrating superior robustness and trustworthiness in real-world fault diagnosis.
Newly built wind farms often face three practical barriers for bearing fault diagnosis: limited fault samples, expensive labeling, and restricted access to historical source-domain data due to privacy and security requirements. This paper presents a source-free domain adaptation framework named dual perturbation robust pseudo-label optimization (DPR-PLO) for cross-domain wind turbine bearing diagnostics using only a pre-trained source model and unlabeled target data. DPR-PLO combines noise-aware hybrid perturbation learning with dynamic prototype alignment to improve robustness against condition-induced distribution shifts. Specifically, it applies time–frequency perturbations and feature-space adversarial perturbations to enhance target diversity, constructs discriminative target representations via dynamic clustering and prototype updating, and adopts a two-stage pseudo-label refinement procedure to reduce noise propagation during self-training. Experiments on Paderborn University Dataset, Jiangnan University Dataset, and a self-collected dataset demonstrate that DPR-PLO achieves 97.56% average accuracy across 18 cross-condition tasks, surpassing benchmark methods by 6.95% and yielding more reliable pseudo-labels. These results indicate that DPR-PLO provides a feasible and robust solution for privacy-preserving bearing health management in new wind turbines.
Haoyang Mao, Yiming Zhang, Weizheng Zhao et al.· Measurement science and tech...· 0 citations
Variations in mechanical load constitute a dominant source of domain shift in data-driven fault diagnosis of rotating machinery, causing models trained under one operating condition to degrade sharply when deployed under another. This work addresses the problem through unsupervised domain adaptation (UDA)—which transfers diagnostic knowledge from a labeled source condition to an unlabeled target condition by aligning their feature distributions—and introduces Padé Approximant Neural Networks (PadéNets) as compact yet highly expressive feature extractors. One-dimensional PadéNet encoders are embedded into three established adaptation frameworks—Deep CORAL, Domain-Adversarial Neural Networks (DANNs), and Conditional Domain-Adversarial Networks (CDANs)—to learn load-invariant representations without any labeled target data. On the Case Western Reserve University benchmark, where the models operate directly on raw time-domain vibration signals, replacing conventional convolutional encoders with PadéNets consistently improves cross-load diagnostic accuracy, reaching up to 99.28% average target-domain accuracy at a low parameter count. To assess generalization to a more demanding setting, the CDAN–PadéNet configuration is further evaluated on frequency-domain representations of the Paderborn University dataset, where domain shift arises from simultaneous variation of load torque and radial force on bearings with real accelerated-lifetime damage, attaining 99.84% average accuracy across six cross-condition transfer tasks while requiring fewer parameters than competing methods. These results establish PadéNet-enhanced UDA as an accurate, broadly applicable approach for robust bearing fault diagnosis under varying operating conditions, with a reduced parameter count suited to resource-constrained embedded platforms.
Sertac Kilickaya, Cansu Celebioglu, Murat Askar et al.· Machines· 0 citations
Deep learning has led to notable progress in fault diagnosis. However, many approaches still rely on the availability of abundant fault samples and on matched train-test distributions, conditions that are seldom met in industrial settings. Although model-agnostic meta-learning (MAML) enables rapid adaptation with limited labeled data, existing fault diagnosis methods based on MAML generally lack explicit modeling of intra-class sample relationships and commonly employ a fixed inner-loop update strategy. To address cross-domain diagnosis with few samples, we enhance MAML by incorporating a Sample Relationship Exploration module that learns intra-class similarity to improve class separability. Specifically, the Sample-level Attention component discovers task-specific affinities, the Explicit Guidance component provides an ideal affinity map to supervise similarity learning, and the Channel-wise Adaptive Fusion component fuses original features with category-aggregated features for classification. In addition, to mitigate overfitting, we replace the fixed inner-loop update scheme in MAML with a trapezoidal gradient descent scheduler that adapts the number of inner-loop update steps across training. Experiments on public and self-collected datasets under multiple cross-domain few-shot settings demonstrate that the proposed method achieves better or more stable diagnostic performance in most evaluated settings.
Zhigang Chen, HaSitieer MaDetihan, Zhihao Zhang et al.· Engineering Research Express· 0 citations
In cross-condition bearing fault diagnosis, the target domain often contains only a subset of the fault classes available in the source domain. This asymmetric label-space setting makes conventional domain alignment prone to negative transfer, because source-private and unreliable samples may still participate in adaptation. To address this problem, this paper proposes a contrastive error-perception-based partial adversarial domain adaptation method, named CEPADA. Instead of relying only on class-level target predictions, CEPADA introduces an error-perception mechanism to jointly evaluate class-level transfer reliability and sample-level uncertainty. Samples with high error risk are assigned lower influence during adversarial alignment, while more reliable shared-class samples are emphasized. Meanwhile, contrastive learning is embedded into the adaptation process to make features of the same fault class more compact and features of different classes more separable. A pseudo-label refinement strategy is also used to provide additional target-domain guidance under the unlabeled setting. Experiments on the case western reserve university and JNU bearing datasets show that CEPADA achieves average accuracies of 93.17% and 86.02%, respectively, with improvements of 1.47 and 1.52 percentage points over the best-performing compared unsupervised baseline in our experiments.
Na Lei, Zhuo Wang, Xue Wang et al.· Engineering Research Express· 0 citations
Cross-domain distribution shifts severely degrade the diagnostic performance of rolling bearing models under unseen variable operating scenarios. Single-source domain generalization (SDG) builds fault diagnosis models using only single-source vibration data, which fits the practical limitations of industrial data collection. Existing contrastive learning methods adopt uniform spectral perturbations for data augmentation, which easily corrupt fault harmonic characteristics and require massive, labeled training samples. To tackle these drawbacks, this paper proposes an unsupervised contrastive learning framework named FDACL. An adaptive frequency-domain augmentation (AFA) module equipped with learnable weights is designed to separate fault-critical frequency bands from noise components. Differentiated amplitude perturbations are applied to two categories of spectral signals to generate diverse pseudo-samples while retaining intrinsic fault information. A shared encoder is trained with combined InfoNCE contrast loss and classification loss to learn domain-invariant fault representations. Validations are carried out on three datasets, namely Case Western Reserve University (CWRU), Paderborn University (PU), and the industrial CRRC Qingdao Sifang railway wheelset bearing dataset acquired from physical test benches. FDACL achieves average cross-speed diagnostic accuracies of 92.68% and 77.85% on CWRU and PU, respectively, and maintains competitive performance on the Qingdao Sifang industrial dataset. It outperforms state-of-the-art baselines by 4.23–8.71% across all SDG transfer tasks. Ablation experiments and hyperparameter analysis verify the efficacy of the AFA module and contrastive learning scheme, providing an unsupervised diagnostic approach for railway bearings under unknown working conditions.
Kai-Sheng Deng, Ping Qu· Italian National Conference...· 0 citations
Existing unsupervised domain adaptation methods for bearing fault diagnosis under varying operating conditions generally assume identical class priors across domains, overlooking the label shift that arises in practice, where healthy-state samples far outnumber fault samples and fault-type frequencies vary with speed, load, and operating history. When feature shift and label shift coexist, prematurely introduced high-intensity alignment can severely degrade minority-class discriminability. We propose DACL-IDA (Dynamic alignment and compactness learning for imbalanced domain adaptation), built on an alignment-scheduling principle rather than a new alignment loss: global adversarial alignment is delayed until classification warmup and black-box shift estimation (BBSE) have stabilized, and is kept restrained thereafter. The BBSE estimate both re-weights the source loss and drives the scheduler, supported by intra-class compactness regularization and long-tailed optimization. A diagnostic experiment separating global adversarial from local class-conditional alignment shows that sustained class-conditional matching (local maximum mean discrepancy) progressively degrades the classifier under severe (reversed-prior) label shift; the final configuration therefore retains only the scheduled global adversarial alignment. On the Paderborn University dataset, under a strictly leakage-free protocol (fixed last-epoch reporting, no target-label model selection; 4 protocols × 12 transfer tasks × 5 seeds), DACL-IDA attains an overall mean balanced accuracy of 89.63% versus 81.30% for the strongest published competitor, DIDA (a deep imbalanced-domain-adaptation framework), winning three of the four protocols by more than 13 percentage points each (paired-t p ⩽ 5.23 × 10−4); on the remaining protocol (I2B), where scheduled alignment is counter-productive, it underperforms both DIDA and the source-only baseline—a failure reported and analyzed rather than omitted. DACL-IDA is also the most convergence-stable method in the stability comparison: outside I2B, its gap between best transient and final-epoch accuracy stays within 2.4 pp, versus 8.2–11.2 pp for DIDA. Cross-dataset experiments on the Case Western Reserve University dataset, a second backbone, imbalance ratios up to ρ = 50, and injected measurement noise support the reported behavior.
Xiangyu Peng, Yang Tao, Lei Hua et al.· Measurement science and tech...· 0 citations