Conformer-PhyFaultNet: Physics-Informed Spectral Attention Conformer for Generalizable Bearing Fault Diagnosis
Rizwan Ullah, Hazrat Bilal, M.S. Aslam, Sarra Ayouni, Abdul Majid, Athanasios V. Vasilakos, Thippa Reddy Gadekallu
Abstract
Intelligent fault diagnosis of rotating machinery is essential for predictive maintenance, yet conventional deep learning models suffer from limited generalization under noisy and cross-domain conditions. Furthermore, their lack of interpretability restricts industrial trust and deployment. To address these challenges, we propose Conformer-PhyFaultNet, a novel physics-informed spectral attention Conformer that seamlessly integrates domain knowledge with advanced sequence modeling. The method embeds characteristic fault frequencies (BPFO, BPFI, FTF) directly into the spectral attention layer, guiding the network toward physically meaningful patterns rather than spurious features. A set of physics-guided tokens is introduced into the Conformer encoder, which persist across layers and acts as stable descriptors of defect signatures. The hybrid spectral attention + physics tokens mechanism enables the model to simultaneously capture local fault harmonics and long-range dependencies across time–frequency representations. Unlike conventional CNN or Transformers, proposed approach ensures interpretability by aligning attention distributions with analytic harmonics and providing layer-wise token activations as diagnostic evidence. This dual mechanism represents the key novelty, bridging analytic fault modeling with modern deep architectures for the first time in a unified framework. Extensive experiments on CWRU, Paderborn, and HUST datasets demonstrate the superiority of the proposed method: in-domain accuracy reaches 93.75, cross-domain transfer achieves 86.75 (CWRU→Paderborn) and 84.3% (HUST→CWRU), while noise robustness remains above 87% at 10 dB and 81% at 0 dB, outperforming CNN, RNN, Transformer and physics-informed baselines by significant margins. The proposed Conformer-PhyFaultNet therefore offers a technically rigorous, interpretable and noise-robust solution that can substantially enhance the reliability and adoption of intelligent predictive maintenance in industrial environments, while also enabling real-time monitoring and edge deployment in industrial IoT systems through low-latency inference (< 5 ms per segment).