Publication

MDO & PHM Lab.

Journal

Physics-guided self-supervised learning for rotating-machinery fault diagnosis: role of average energy density equation
Year
2026
Author
Jaewook Lee, Sukeun Hong, Sanghoon Lee, Jongsoo Lee
Journal
Advanced Engineering Informatics
Vol.
69
Issue Date
2026.01
Self-supervised learning (SSL) has recently garnered significant attention in the fault diagnosis of rotating machinery for reducing data-processing costs and achieving effective feature representation learning without relying on data labeling. However, existing SSL methods depend significantly on augmentation strategies and contrastive loss. Recently, efforts have been undertaken to overcome augmentation and contrastive loss limitations. Additionally, generalizable SSL models that can be utilized in various domains are being actively investigated. Hence, this study proposes a physics-guided self-supervised learning (PgSSL) framework that can be generalized and utilized for engineering purposes, such as fault diagnosis. This method uses spectrograms obtained from the vibration and acoustic signals of rotating machinery to apply the average energy density equation such that the learning model can be trained to output physically meaningful representations in the feature space. Through equation-based learning, domain scalability is increased. This enables representation learning under various operating conditions of rotating machinery, including unseen domains, and fault diagnosis with limited or imbalance data. By performing various experiments on bearing faults, we verified that the proposed PgSSL framework is robust, generalizable, and efficient for the fault diagnosis of various rotating machines. Based on the experimental results, the proposed PgSSL achieved an accuracy exceeding 0.9645 and an F1-score above 0.9637, even under imbalanced conditions, thus demonstrating its strong robustness in actual industrial scenarios. Furthermore, the proposed method consistently outperformed conventional contrastive SSL and lightweight transformer models for balanced and imbalanced configurations.