Paper: Toward Regulation-Friendly Bearing Fault Diagnosis via Structural Embedding Preserving Machine Unlearning
Authors: Jiyoung Lee (CAU undergraduate researcher), Chaewon Yoon (CAU undergraduate researcher), Hoki Kim (CAU)
Venue: Advanced Engineering Informatics (SCIE top ~5%)
Our lab paper “Toward Regulation-Friendly Bearing Fault Diagnosis via Structural Embedding Preserving Machine Unlearning” has been published in Advanced Engineering Informatics, a journal in the top ~5% of SCIE. Jiyoung Lee (first author) and Chaewon Yoon carried out the work as undergraduate researchers in our lab.
Research
Bearing fault diagnosis supports predictive maintenance by detecting mechanical failures early. In practice, some training data may need to be removed because of privacy regulations, contractual obligations, or data quality issues. Machine unlearning removes the influence of specific data without retraining the model from scratch.
The paper first analyzes existing unlearning methods on fault diagnosis systems and shows that they perform poorly on the time-series vibration data typical of bearing faults. To address this, it proposes Soft-directed Embedding Unlearning (SEU), which manipulates the embedding space directly: it pushes forget samples away from their original class clusters while preserving the structure of retain samples using soft examples near decision boundaries.
Measured with a manifold similarity score, SEU stably approximates the ideal manifold structure, and on both public and private industrial data it achieves more effective unlearning than the evaluated baselines.
Related
- Earlier work by the same team: Published in IEEE TII — Adversarial Retain-Free Unlearning