HYBRID DEEP LEARNING MODELS FOR EQUIPMENT FAILURE PREDICTION IN U.S. INDUSTRIAL SYSTEMS

Abstract

Equipment failure prediction plays an important role in U.S. industrial systems, where unexpected downtime carries heavy economic and safety consequences. Deep learning methods are now widely promoted for predictive maintenance, yet empirical evidence showing that hybrid architectures consistently outperform strong classical models under realistic, leakage-safe evaluation remains limited. This study examines early failure prediction using the NASA C-MAPSS turbofan engine dataset and frames the task as a binary classification problem that flags failure-imminent conditions within a fixed prediction horizon. The analysis relies on a rigorous experimental pipeline that includes engine-level data partitioning, sliding-window temporal representations, and carefully defined failure labels. Classical baseline models built on engineered statistical features are evaluated alongside several deep learning architectures, including LSTM, CNN, CNN–LSTM hybrids, and LSTM models enhanced with attention mechanisms. A series of ablation studies explores the practical value of architectural hybridity, the influence of temporal window length and prediction horizon, and sensitivity to sensor removal. The results show that classical models, with gradient boosting as a notable example, deliver very strong performance with a healthy balance between precision and recall. Deep learning models reach comparably high ROC-AUC values, yet their recall for imminent failures drops sharply under standard decision thresholds. The ablation findings further reveal that hybrid architectures do not consistently outperform simpler designs and that performance depends strongly on temporal configuration and sensor choice. Taken together, these results indicate that hybrid deep learning models do not automatically earn their added complexity for equipment failure prediction. The study reinforces the value of strong baselines, transparent evaluation practices, and decision-oriented metrics in predictive maintenance research.