acceptodds
Under review as a conference paper at ICLR 2027

Class-Conditional Progressive Calibration under Selective Decoupled Domain Adaptation for Long-Tailed Cross-Condition Bearing Fault Diagnosis

Abstract

Rolling bearings are critical supporting components of rotating machinery and directly affect the safe and reliable operation of industrial equipment. Rolling bearing fault diagnosis remains challenging when operating-condition variations coexist with long-tailed fault distributions. Under such settings, uniform domain alignment may introduce unreliable transfer across operating conditions, while long-tailed class priors can bias target supervision toward head classes. To address these issues, we propose Selective Decoupled Domain Adaptation with Class-Conditional Progressive Calibration (SDDA-CCPC). Specifically, SDDA separates fault-discriminative and condition-related representations and performs reliability-weighted domain alignment to reduce the influence of unreliable cross-condition transfer. CCPC calibrates target predictions under long-tailed class priors and combines classifier–prototype agreement with progressive within-class selection to construct more reliable target supervision for minority classes. Across six long-tailed unsupervised cross-condition transfer tasks under different load conditions, SDDA-CCPC achieves an average Macro-F1 of 0.9178 and improves both overall diagnostic performance and tail-class recognition. For the extreme tail inner-race fault class, the average F1-score improves by 3.91%. Ablation results further demonstrate the complementary contributions of SDDA and CCPC.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
rolling bearing fault diagnosis, cross-condition fault diagnosis, long-tailed learning, unsupervised domain adaptation, class-conditional reliability calibration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_h61hOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors introduce a solution for the cross-domain bearing fault diagnosis under class imbalance. They introduce two modules. SDDA is proposed to separate adaptation features from those of classification via orthogonality and adversarial alignment constraints. To improve the supervision on minority class, they then introduce CCPC to combine the source-prior-based prediction correction, classifier-prototype agreement and within-class pseudo-label selection. The empirical evaluation is conducted on 6 load-transfer tasks from the CWRU dataset with a manually set class-imbalance ratio.

Critical strengths
  1. The proposed class-aware pseudo-label selection is interesting. Using a classifier-prototype agreement with within-class ranking for accepting minority-class samples is a reasonable mechanism to handle class-imbalance problem.

  2. The authors introduce a manually setting for developing class imbalance problem using the bearing fault diagnosis benchmark dataset (CWRU).

  3. It is also interesting that they chronologically split the raw vibration sequences before making windows, which reduces the leakage risk caused by randomly splitting from overlapping windows.

Critical weaknesses
  1. My key concern is formulation of the problem and the limited contributions. The proposed components are all generic for class imbalanced domain adaptation problem, without any specific component designed for bear fault signals. The class imbalance problem is a well-studied problem in the domain adaptation field. Thus, it is important to also conduct evaluation on the benchmark datasets for class imbalanced domain adaptation. At least, it is expected to present results Office-Home RS-UT and imbalanced DomainNet during rebuttal. Meanwhile, the latest baselines (at least after 2025) should be included as baselines for comparisons.

  2. A followup concern is the novelty and contribution of the proposed work. The stated challenges, including domain shift, head-class bias and unreliable minority-class pseudo-labels, are all well-studied problems in domain adaptation. The proposed method likewise combine standard feature separation, confidence weighting, prior correction and class-aware self-training, without exploring the characteristics that are unique for bearing fault signals.

  3. The authors manually choose a 40:20:10:4:8:1 class distribution on CWRU without any statistical significance analysis or concrete evidence to support this ordering. Normal-versus-fault imbalance does not justify the assumed long tail across fault types and severities. Moreover, vibration-window counts do not represent independent fault occurrences.

  4. The implementation details are missing. The reported results lack statistical significance. I am also wondering the motivation of using the unusually short 20-point windows (approximately 1.67 ms). Please provide justification and sensitivity analysis for this window size.

  5. The overview figure (Figure 2) is too unprofessional. Please consider re-draw Figure 2 with a better design.

Questions for authors

See my concerns in Critical Weaknesses.

3 comments

We thank the reviewer for the detailed comments on the methodological contributions, experimental design, and presentation. We have strengthened the evaluation, clarified the experimental protocol, expanded the implementation details, and revised the framework illustration. Our responses are provided below.

  1. Evaluation benchmarks and recent baselines:

To extend the evaluation beyond CWRU while retaining the vibration-based fault-diagnosis setting, we have added experiments on HUST. The two datasets differ in operating conditions, fault categories, sampling frequencies, window lengths, and class-imbalance configurations. As reported in Appendix F, SDDA-CCPC achieves the highest six-task average Macro-F1 of 0.9343, Accuracy of 0.9467, and Macro-Recall of 0.9511 among the evaluated methods. We have also incorporated two recent methods, CIDSAN (2025) and DMTVCD (2026), into the baseline comparisons. These additions extend the empirical evaluation across two bearing datasets and provide comparisons with recent approaches to class-imbalanced cross-condition fault diagnosis.

  1. Bearing-signal characteristics and methodological contributions:

The methodological design of SDDA-CCPC is closely related to the characteristics of vibration-based bearing diagnosis. Changes in operating loads can alter vibration responses, while different fault locations and defect sizes produce distinguishable local signal patterns. The challenge is therefore to preserve fault-relevant information during cross-condition adaptation while maintaining reliable recognition of sparsely represented fault categories. These considerations are reflected in two aspects of our design. First, the input representation incorporates raw-waveform, local morphological, and FFT-based circular autocorrelation features to characterize local vibration patterns. Building on these representations, SDDA jointly controls the feature space subjected to domain alignment and the contribution of individual samples, reducing the influence of condition-dependent variations on fault discrimination. Second, CCPC addresses the difficulty of identifying reliable target samples across fault locations and severities with highly unequal sample frequencies. Its calibration, prototype verification, and within-class selection form a dependent process that controls pseudo-label reliability and reduces head-class dominance during target supervision. The contribution thus lies in translating these diagnostic challenges into specific representation-alignment and target-supervision mechanisms, connecting vibration-signal characteristics with reliability-aware cross-condition adaptation.

  1. Construction of the class-imbalanced distribution:

We construct the 40:20:10:4:8:1 class distribution on CWRU to establish a controlled evaluation setting that examines cross-condition adaptation under multiple levels of sample scarcity. The design goes beyond a simple Normal-versus-Fault imbalance. It includes a dominant Normal class, several moderately represented fault classes, a moderately under-represented 14 mil outer-race fault, and an extremely under-represented 21 mil inner-race fault. This arrangement allows us to evaluate fault recognition across different fault types, defect sizes, and degrees of class imbalance, particularly whether minority classes remain distinguishable under operating-condition shifts. Importantly, the same class proportions are maintained across all four operating conditions. This keeps the imposed imbalance configuration consistent across transfer tasks, allowing performance differences to be examined under different operating-condition shifts. The specific allocation defines experimental sampling proportions rather than physical fault occurrence frequencies. The reported counts represent sampled vibration windows, not independent fault events. We have clarified this construction in Appendix B.2. Additionally, the HUST experiments in Appendix F employ a different 10:1 imbalance configuration, extending the evaluation to another class-distribution setting.

AuthorsOpenReview ↗

Response 2

  1. Implementation details, statistical reliability, and window length:

We have expanded Appendices B and G to provide detailed information on data partitioning, network configuration, optimization settings, and hyperparameters. Appendix G.2 also presents a sensitivity analysis of four key SDDA-CCPC hyperparameters. For statistical reliability, Table 3 now reports the mean and standard deviation across five independent random seeds (42–46) on all six CWRU transfer tasks. SDDA-CCPC achieves an overall Macro-F1 of 0.9206 ± 0.0027. Appendix G.3 (Tables 16–17) provides the complete seed-wise results and additional statistics for the extreme tail class. The CWRU experiments use a fixed 20-point input window, corresponding to approximately 1.67 ms at 12 kHz. This configuration supports classification through learned local vibration patterns. Appendix B.1 and B.5 provide the chronological data partition procedure, input tensor dimensions, and feature extraction details, including the raw-waveform, local morphological, and FFT-based circular autocorrelation representations.

  1. Presentation of Figure 2:

We have redesigned Figure 2 to provide a clearer representation of the SDDA-CCPC framework. The revised illustration distinguishes the SDDA and CCPC components and explicitly depicts source prototype memory, classifier–prototype agreement, and progressive within-class reliability ranking. The information flow and module relationships have also been refined to improve readability and highlight the complementary functions of the two proposed mechanisms.

Thanks the authors for the detailed response. Most of my concerns are properly addressed. I am happy to maintain my positive rating.

Reviewer_1pw2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of unsupervised cross-condition rolling bearing fault diagnosis under severe class imbalance (long-tailed distribution). The authors propose SDDA-CCPC (Selective Decoupled Domain Adaptation with Class-Conditional Progressive Calibration). The framework consists of two main modules:

Selective Decoupled Domain Adaptation (SDDA): Decouples latent representations into a fault-discriminative branch fcf_c and a condition-adaptation branch fdf_d via an orthogonality constraint, while weighting the adversarial domain discriminator using prediction certainty and ground-truth confidence.

Class-Conditional Progressive Calibration (CCPC): Calibrates target predictions using moving estimates of class priors, enforces agreement between the classifier output and source prototype memory, and progressively selects target pseudo-labels via intra-class relative ranking.

The method is evaluated on the Case Western Reserve University (CWRU) bearing dataset across 6 load-transfer tasks under a 40:1 class imbalance ratio, showing an average Macro-F1 improvement from 0.8991 (IRDDDA) to 0.9178.

Critical strengths
  1. Systematic formulation of coupled challenges: The paper effectively isolates two distinct failure modes in transfer diagnosis: non-uniform domain transferability at the sample level, and confirmation bias toward majority classes during pseudo-labeling under long-tailed distributions.

  2. Structured target pseudo-supervision pipeline: The combination of dual-view agreement (classifier prediction vs. feature-to-prototype cosine similarity) and progressive intra-class rank selection R_k^{(e)} represents a sound approach to prevent tail classes from being suppressed by global confidence thresholds.

Critical weaknesses
  1. At a 12 kHz sampling rate and ~1750 RPM, a 20-point window spans only 1.67 ms (≈1/20\approx 1/20th of a shaft rotation), whereas characteristic bearing fault periods require at least 80–150 points. Capturing periodic impact kinematics or fault frequencies from such short slices is physically impossible, indicating either a severe typographic error or an invalid experimental setup.

  2. Standard Bayesian adjustment requires dividing by the source prior pispi_{s} to counteract head-class bias learned from an imbalanced training set. Placing pis,kpi_{s,k} in the numerator of Eq. (5) does the opposite, mathematically penalizing minority classes further and exacerbating majority-class bias.

  3. The evaluation relies solely on the saturated CWRU benchmark and reports isolated point estimates without standard deviations over multiple random seeds. Validating against additional diagnostic datasets and modern baselines with statistical significance tests is necessary to support the claimed gains and improve the score.

Questions for authors

1.Please clarify whether L=20L=20 is a typo, as an isolated 20-point slice at 12 kHz cannot physically capture bearing fault periods that require 80–400 points. If an unmentioned transform (e.g., FFT or CWT) was applied prior to feature extraction, please specify the exact preprocessing steps and tensor shapes.

  1. Please provide a formal Bayesian derivation justifying why the source prior \mathrm{\Pi}_{s,k} is in the numerator, as standard logit adjustment requires an inverse relationship to de-bias predictions against head classes.

  2. Could the authors report the mean and standard deviation across at least 5 independent random seeds for Table 3 to confirm the statistical reliability of the results?

3 comments

We thank the reviewer for the detailed comments. We have clarified the input representation and preprocessing procedure, expanded the mathematical justification of prior calibration, and strengthened the experimental evaluation through additional data, recent baselines, and multi-seed results. Our responses are as follows.

  1. Clarification of the 20-point input window and preprocessing:

The 20-point window is the actual input configuration used in the CWRU experiments, corresponding to approximately 1.67 ms at a sampling frequency of 12 kHz. The model learns discriminative local vibration patterns from these short segments for fault classification, rather than estimating complete fault-impact periods or characteristic fault frequencies. We have expanded Appendix B.1 and B.5 to clarify the preprocessing procedure and network architecture. The raw vibration signals are first divided chronologically into 80% training and 20% test segments, followed by window generation within each partition. The input tensor is reshaped from B × 20 to B × 1 × 20 for one-dimensional convolution. The backbone produces an intermediate representation of B × 96 × 20, which is subsequently projected into a 64-dimensional latent feature. The multi-view configurations incorporate raw-waveform information, local morphological characteristics, and normalized circular autocorrelation computed through FFT-based operations within the original 20-point window. These details establish the relationship between the input configuration, feature extraction, and the classification objective. The supplementary HUST experiments further evaluate SDDA-CCPC using 256-point vibration windows at 25.6 kHz, corresponding to a 10 ms input duration, under different operating conditions and fault categories.

  1. Mathematical justification of the prior calibration mechanism:

We have expanded Appendix A.2 to explain why the source prior appears in the numerator of Eq. (5). The proposed method uses the source prior as a reference for calibrating target predictions. This differs from standard Bayesian label-shift correction, which adjusts a source posterior using the target-to-source prior ratio. Specifically, our calibration factor is defined as the source-class prior divided by the estimated target prediction proportion. Appendix A.2 shows that the resulting adjustment is the closed-form solution of a KL-regularized optimization objective. The derived pairwise log-probability relation further shows that classes overrepresented in target predictions relative to the source reference receive lower relative probabilities. This adjustment is applied before classifier–prototype verification and class-wise target-sample selection.

  1. Additional evaluation and multi-seed reliability:

We have strengthened the experimental evaluation in three aspects. First, Table 3 now reports the mean and standard deviation across five independent random seeds (42–46) for all six CWRU transfer tasks. SDDA-CCPC achieves an overall Macro-F1 of 0.9206 ± 0.0027. Appendix G.3 (Tables 16–17) provides the complete seed-wise results and additional statistics for the extreme tail class. Second, we have incorporated two recent class-imbalanced bearing fault diagnosis methods, CIDSAN (2025) and DMTVCD (2026), into the baseline comparisons. Third, Appendix F presents additional experiments on the HUST bearing dataset, covering six cross-condition transfer tasks under a 10:1 imbalance ratio. SDDA-CCPC achieves the highest six-task average Macro-F1 of 0.9343, average Accuracy of 0.9467, and Macro-Recall of 0.9511 among the compared methods. These additional evaluations, together with the five-seed CWRU results, provide broader empirical evidence of the proposed method's diagnostic performance across operating conditions and its recognition capability for under-represented fault classes.

Thank you for the detailed response and additional experiments. These revisions have addressed some of my concerns. However, I would still appreciate more direct experimental evidence supporting the choice of a 20-point window. The sensitivity analysis in Table 15 does not cover window length, and the 256-point window experiments on HUST cannot replace a controlled comparison of different window lengths on the same dataset, since the dataset, sampling rate, and other settings also differ. Could the authors compare the proposed method with the main baselines using different window lengths on CWRU, while keeping the other settings as consistent as possible? This would help clarify whether the reported performance gains depend on the specific choice of a 20-point window.

Thank you for your suggestion. We have conducted additional experiments on the CWRU dataset using window lengths of 64 and 128 points, with the original 20-point setting included for comparison. Three cross-condition transfer tasks were considered, and the proposed method was compared with the main baselines.

The experimental settings and results are provided in the newly added Appendix H (Tables 18–20). We sincerely appreciate your valuable suggestion, which has helped us strengthen the experimental evaluation.

Reviewer_WHDdOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SDDA and CCPC to solve two major problems in the adaptive bearing fault diagnosis method, including negative migration caused by different sample alignment during condition migration and category bias of pseudo label confidence under long tail distribution, which effectively improves the diagnosis performance. Combined with a large number of experimental results, the effectiveness of the method is fully proved.

Critical strengths

The paper proposes Selective Decoupled Domain Adaptation with Class-Conditional Progressive Calibration (SDDA-CCPC). Specifically, SDDA separates fault-discriminative and condition-related representations and performs reliability-weighted domain alignment to reduce the influence of unreliable crosscondition transfer. CCPC calibrates target predictions under long-tailed class priors and combines classifier–prototype agreement with progressive within-class selection to construct more reliable target supervision for minority classes. Across six long-tailed unsupervised cross-condition transfer tasks under different load conditions, SDDA-CCPC achieves an average Macro-F1 of 0.9178 and improves both overall diagnostic performance and tail-class recognition. For the extreme tail inner-race fault class, the average F1-score improves by 3.91%. Ablation results further demonstrate the complementary contributions of SDDA and CCPC.

The innovation is outstanding, and the difficulty of bearing fault diagnosis is effectively solved. The writing is clear and readable. The experiment is sufficient, which can obviously reflect the advantages of innovation.

Critical weaknesses
  1. The design idea of the designed module is similar to that of the existing methods, but it needs to be further clarified.

  2. The method in this paper contains several hyper parameters, such as γ\gamma, but lacks sensitivity analysis of hyper parameters.

  3. The explanation of variables in the formula is not detailed enough.

Questions for authors
  1. The design idea of the designed module is similar to that of the existing methods, but it needs to be further clarified.

  2. The method in this paper contains several hyper parameters, such as γ\gamma, but lacks sensitivity analysis of hyper parameters.

  3. The explanation of variables in the formula is not detailed enough.

2 comments

We thank the reviewer for the constructive comments. We have strengthened the presentation of the methodological contributions, added hyperparameter sensitivity analysis, and expanded the mathematical explanations in the appendix. Our detailed responses are provided below.

  1. Clarification of the methodological contributions:

The methodological contribution of SDDA-CCPC is centered on two specific changes to the adaptation process: reliability-constrained representation alignment and class-conditional target supervision.For SDDA, the key insight is that controlling alignment strength alone is insufficient when operating-condition information and fault-discriminative information remain entangled. SDDA therefore jointly constrains where alignment is performed and how strongly each sample participates, restricting reliability-weighted adversarial learning to a condition-related representation while maintaining a separate fault-discriminative space.For CCPC, the central issue is that prediction confidence and pseudo-label correctness become increasingly difficult to distinguish under long-tailed class distributions. Class-wise selection can retain erroneous minority-class predictions, while confidence verification can preserve head-class preference. CCPC addresses these issues by linking prior calibration, classifier–prototype verification, and within-class reliability assessment in a sequential process. Each stage establishes the conditions under which the next stage operates.These designs establish two distinct reliability controls within the adaptation process: one governs representation transfer, and the other governs the admission of target pseudo-supervision. Together, these two mechanisms allow SDDA-CCPC to control representation alignment and target supervision using different reliability criteria.

  1. Hyperparameter sensitivity:

We have added a hyperparameter sensitivity analysis in Appendix G.2 (Table 15), examining four key parameters: the adversarial alignment weight, orthogonality weight, training-time prior calibration coefficient, and maximum sample retention ratio.The analysis uses a one-factor-at-a-time evaluation on the CWRU T0→T3 transfer task. Across eight parameter perturbations, Macro-F1 ranges from 0.9011 to 0.9107. The orthogonality weight exhibits the largest difference between its evaluated values, whereas the training-time prior calibration coefficient shows the smallest variation.The corresponding default parameter configuration is summarized in Appendix G.1 (Table 14).

  1. Explanation of mathematical notation:

We have expanded Appendix A to provide detailed definitions of the mathematical notation and implementation procedures, including the mini-batch notation in orthogonality regularization, the target-prior tracker and its momentum update, source prototype construction, class-conditional reliability ranking weights, and the terms involved in pseudo-label learning and prototype alignment. We have also supplemented the mathematical interpretation of prior calibration through its optimization formulation and the resulting adjustment of relative class probabilities. These additions provide a clearer description of the proposed mechanisms and their optimization procedures.

Thank you for the detailed response. The additional explanation of the method has helped me better understand the different roles of SDDA and CCPC.

The added sensitivity analysis addresses my questions about the four key hyperparameters. The additional explanations of the mathematical notation and implementation in Appendix A also help address my previous concern that the variables in the equations were not explained in sufficient detail.

I am happy to maintain my rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.