Remaining Useful Life Prediction of Bearings via Semi-Supervised Transfer Learning Based on an Anti-Self-Healing Health Indicator

Remaining useful life (RUL) estimation of a bearing is a methodology to monitor rolling bearings for a system’s performance and reliability. It predicts the exact residual time without operational interruptions until complete bearing failure by training a deep learning model to predict the remaining time of working using extracted signal features. Extracting features is one of the most important subjects since its quality directly influences the performance of predicting RUL. Features should gradually and consistently increase over time and capture sudden deterioration within normalized specific thresholds. However, recent studies have not addressed feature extraction methods that consider all of these aspects. Moreover, some bearings exhibit a “self-healing” phenomenon, in which bearing conditions appear to temporarily improve, and this complicates the accurate representation of consistent performance degradation. However, very few studies have properly addressed this issue. Meanwhile, transfer learning is frequently used when training the RUL deep learning model because there is a lack of data for run-to-failure experiments. Most RUL estimation methodologies pre-train and apply deep learning models with supervised learning. But supervised transfer learning supposes that researchers already have access to end-of-life (EOL) data—often unavailable in industrial settings—limiting their practicality. To address these challenges, this paper proposes a novel semi-supervised transfer learning methodology that integrates an anti-self-healing health indicator (ASH-HI) with a transformer-based architecture. ASH-HI is a health indicator that quantifies the power spectrum density (PSD) difference between normal and abnormal states using skewness-based parameter selection, eliminating the need for manual parameter tuning. Also, it overcomes the self-healing problem by measuring the difference not only between normal and abnormal states but also between “correction” and abnormal states. Also, this paper presents a new semi-supervised transfer learning method without EOL information. The proposed methodology is validated using the PHM 2012, NASA IMS, and an experimental setup. This study is the first to attempt transfer learning using more than three datasets simultaneously, resulting in significantly improved performance.