PSI - Issue 80

Nima Rezazadeh et al. / Procedia Structural Integrity 80 (2026) 411–417 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

412

2

for detecting fault signatures, outperforming acoustic and thermal methods by directly revealing dynamic subsurface behavior (Zhang et al., 2022). Machine-learning and deep-learning methods have significantly improved fault detection but still face challenges. These include limited labeled data, domain shift across operating conditions, and a lack of interpretability that reduces technician trust (Rezazadeh et al., 2025, 2024). Existing domain adaptation strategies such as model-based fine-tuning, feature alignment, instance reweighting, and prototype alignment often suffer from overfitting, uniform treatment of data, or poor transparency in how target samples contribute to learning. To address these limitations, this study introduces Proto-ADDA-X (Prototype-Attention-based Adversarial Domain Adaptation with Explainability). This framework aligns class prototypes between domains and assigns weights to target samples based on similarity and prediction confidence. It enables accurate bearing fault diagnosis using minimal labeled data and low computational cost. Performance is validated on two benchmark datasets under different speeds and loads. The framework also provides interpretability through similarity maps and feature visualizations that highlight the vibration patterns influencing each decision. Early approaches to address domain shift in bearing diagnosis used fine-tuning on limited target data, which improved accuracy under moderate shifts but often led to overfitting when conditions varied significantly (Asutkar and Tallur, 2023). To mitigate such mismatches more broadly, researchers developed feature-alignment techniques that match source and target distributions using methods like maximum-mean-discrepancy and adversarial learning, including domain adversarial networks and contrastive or multi-adversarial variants (Pan et al., 2025). Instance reweighting methods followed, adjusting sample weights dynamically to reduce negative transfer in imbalanced or noisy settings, as demonstrated in deep-imbalanced and dynamic-reweighted schemes (Ding et al., 2023). Prototype based strategies were later proposed to align class centroids across domains, either using fixed or adaptively updated prototypes, improving class separation across conditions but offering limited clarity on which samples shape each prototype (Xie et al., 2025). To improve interpretability, post-hoc explainable-AI tools have been applied to black-box models. Techniques like Grad-CAM highlight key time-frequency regions in vibration spectrograms (Kim and Kim, 2024), while layer-wise relevance propagation provides detailed pixel-level attributions aligned with physical signal behavior (Herwig and Borghesani, 2023). Integrated gradients reveal root-cause frequencies by tracing input relevance paths (Peng et al., 2022), and SHAP-based rankings identify influential features to support feature selection and build user trust (Magadán et al., 2024). However, these methods do not influence training directly, leaving a gap in integrating interpretability with domain adaptation. Most domain adaptation studies for bearing diagnosis reduce global distribution shifts or use fixed prototypes, however they still misalign rare fault classes, add little interpretability, and often require heavy adversarial computations. Post-hoc explainability tools reveal important signal regions but do not influence how transferable features are learned. The proposed prototype-attention framework instead updates class centroids and sample weights jointly during training, producing built-in attention maps that highlight fault-critical vibration cues while keeping the model lightweight and adaptable across operating conditions. 2. Methodology 2.1. Feature extraction Although deep learning models can automatically learn features from the raw vibration signal ( ) , this study employs manual feature engineering to focus on the design of the domain adaptation and explainability modules. To this end, each ( ) is represented by a concise vector of interpretable descriptors (root‐mean‐square, kurtosis, crest factor, dominant frequency, spectral centroid, mean MFCC coefficients, and discrete wavelet sub‐band energies) . These features are standard in vibration‐ba sed bearing diagnosis and provide clear inputs for robust transfer and transparent prototype‐attention mapping (Li et al., 2022).