Fault-Aware RUL Prediction for Diesel Engines Using Physics-Motivated Degradation Models

Abstract

Remaining Useful Life (RUL) prediction is a key problem in Prognostics and Health Management for diesel engines operating under diverse fault conditions. Although data driven models such as Long Short Term Memory (LSTM) networks have shown strong predictive capability, their performance differs greatly depending on the type of fault. This paper investigates influence of fault specific degradation physics on RUL predictability. This is achieved by modelling six common diesel engine faults using physics motivated degradation formulations and synthetic run-to-failure datasets. This work involves training separate LSTM models for each fault. Separate LSTM models has been trained for each fault under a strictly leakage-free evaluation protocol. The paper further demonstrates that attainable RUL accuracy is governed primarily by degradation time scale and signal observability rather than model architecture alone. The proposed framework is intended for controlled comparative analysis rather than direct field deployment, and it highlights the need for fault-aware evaluation in data-driven prognostics. © 2026 IEEE.