Decomposed Degradation Pattern Alignment for Domain Adaptation in Machine Remaining Useful Life Prediction

In complex manufacturing processes, multiple sensors have been installed on the machine, enabling the data-driven Remaining Useful Life (RUL) prediction. Timely RUL predictions reduce costs, minimize idle time, and prevent cascading failures in interconnected systems. While high-performance deep learning models have improved RUL prediction accuracy, domain shifts caused by changes in equipment or operating environments often lead to prediction failures. Domain adaptation techniques have been introduced to address this issue, yet they often struggle to align the degradation patterns embedded in run-to-failure data when extracting domain-invariant features. This study presents a novel approach that decomposes time-series sensor data into trend and periodic components, extracts adapted representations for each, and effectively aligns them to improve cross-domain RUL prediction performance. Our proposed method demonstrates superior RUL prediction performance across 12 cross-domain scenarios of the C-MAPSS dataset compared to four baseline models. Notably, our approach proves to be robust and effective in accurately predicting RUL, even in scenarios with significant distributional differences between domains.