Multi-class Similarity-based Approach for Remaining Useful Life Estimation

Estimating the Remaining Useful Life (RUL) allows for effective machine maintenance and cost reduction. Lately, RUL estimation has shifted to data-driven methods, emphasizing autoencoder (AE) architectures and similarity-based metrics. However, current approaches often overlook failure modes in constructing health index (HI) libraries, impacting accuracy. This work proposes a RUL estimator using AE and similarity, explicitly considering failure modes. It adopts a multi-class approach, categorizing processing paths based on distinct failure modes. An online classifier identifies the failure mode from input data, guiding the RUL prediction phase for more accurate estimates. This approach enhances accuracy by tailoring data processing to the failure class, capturing patterns for precise estimation. Experiments using the C-MAPSS aircraft engine simulator dataset compare the proposed scheme with a reference case ignoring failure classes, emphasizing the significance of considering distinct failure modes in RUL estimation.