Integrating Signal Decomposition, Stochastic Modeling, and Deep Learning for Interpretable Predictive Maintenance

This paper proposes a novel hybrid framework for predictive maintenance that integrates Singular Value Decomposition (SVD), Exponential GARCH (E-GARCH) modeling, and Long Short-Term Memory (LSTM) networks. The framework decomposes multivariate sensor signals into nominal and anomalous components using windowed SVD, models the volatility of residuals through E-GARCH, and inputs volatilityweighted signals into an LSTM network to support various maintenance objectives such as remaining useful life (RUL) prediction and anomaly detection. The main contribution of this work is the development and validation of a volatility-aware prognostic framework that combines physical signal decomposition, probabilistic modeling of anomaly uncertainty, and deep sequential learning. This hybrid design offers interpretable and computationally efficient estimations. The proposed approach is evaluated on the NASA C-MAPSS FD001 dataset, demonstrating a lower RMSE (14.22) compared to a standard LSTM baseline (16.14), along with improved early-warning capabilities. While it may not outperform deep black-box models in terms of raw predictive accuracy, the method's enhanced interpretability and computational efficiency make it a compelling choice for real-time, safety-critical applications requiring transparent and reliable prognostics.