Multi-Horizon Engine Fault Prediction Using Lightweight Deep Learning and Transformer Models

Predictive maintenance is critical in modern automotive system operations as vehicles increasingly rely on complex sensor networks and electronic control units. Traditional fault detection systems detect failures close to the fault event, leaving minimal reaction time for safe intervention. This work is an attempt in this direction for enabling early and reliable fault prediction, several timesteps before the actual malfunction. Presents a lightweight and explainable deep learning framework for multi-horizon lead-time prediction, integrating a unique lead time evaluation pipeline. Multiple architectures—CNN, BiLSTM, Transformer, and hybrid combinations—are evaluated on EngineFaultDB dataset using a leakage-free, run-aware windowing pipeline that preserves temporal integrity. Experiments reveal two key precursor zones: a short-term and a long-term degradation zone. The proposed CNN–BiLSTM hybrid achieves an F1-score of 0.80 for long-horizon prediction.To ensure inter-pretability and trust, the system integrates SHAP, highlighting the contribution of critical sensors—across temporal windows that precede the fault event. Overall, the approach provides an accurate, interpretable, and computationally efficient solution for proactive fault forecasting in vehicle platforms.