Predictive Maintenance on C-MAPSS Using LSTM Variants and Attention

Accurate prediction of Remaining Useful Life (RUL) is critical for effective predictive maintenance in industrial systems. This study improves RUL forecasting performance through signal processing techniques, avoiding the complexity of hybrid deep learning models. Using the NASA Commercial Modular Aero-Propulsion System Simulation FD001 dataset, which provides run-to-failure sensor data from turbofan engines, we applied Kalman Filtering and multi-level Discrete Wavelet Transform for noise reduction and feature enhancement. Low-variance features were discarded, and Min-Max normalization was used for data scaling. Input sequences were generated using a sliding window of length 50. Four Long Short-Term Memory (LSTM)-based models were developed for comparison: a baseline LSTM, a Bi-directional LSTM, and two variants incorporating Multi-Head Attention. While attention mechanisms improved average performance in terms of Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) by enhancing temporal focus, they also introduced higher variance and occasional large errors. In contrast, signal processing significantly improved input quality, model stability, and convergence, enabling simpler architectures to achieve competitive results. All models were trained using MSE loss and evaluated on MAE and RMSE. The results highlight that well-designed signal processing pipelines can enhance RUL prediction accuracy while reducing reliance on complex neural network architectures.