Adaptive Temporal Feature Learning for Early Fault Prediction in Industrial Equipment under Changing Operating Conditions

Abstract

Early fault prediction is essential for improving equipment reliability, reducing unplanned downtime, and optimizing maintenance strategies in industrial environments. However, conventional machine learning approaches often struggle to maintain consistent prediction performance under changing operating conditions, noisy sensor measurements, and limited fault observations. This study proposes an adaptive temporal feature learning framework for early fault prediction in industrial equipment. The proposed framework integrates multi-scale temporal feature extraction, operating-condition-aware attention, and uncertainty-guided feature weighting to capture evolving degradation patterns across heterogeneous sensor streams. To improve generalization across operating regimes, a lightweight distribution alignment mechanism is introduced to reduce feature discrepancies between different operating conditions. The framework is designed for evaluation using the NASA C-MAPSS turbofan engine degradation dataset, which provides multivariate sensor measurements across multiple simulated engine operating scenarios. Experimental evaluation will compare the proposed approach against Random Forest, XGBoost, LSTM, and Transformer-based prediction models. Performance will be assessed using remaining useful life prediction errors, including RMSE and MAE, together with early-warning performance under different prediction horizons. Additional experiments will investigate robustness to sensor noise, missing measurements, and distribution shifts. The proposed framework aims to provide reliable degradation forecasting and support more efficient condition-based maintenance decisions in industrial systems.