Machine Learning-Based Remaining Useful Life Predictions and Its Application on Predictive Maintenance

Predictive maintenance marks a significant improvement over traditional methods by preventing failures, minimizing downtime, and reducing costs, especially in aviation where it enhances safety. This paper explores the impact of window size on Remaining Useful Life (RUL) estimation using the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) dataset. The performance of Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) is evaluated in handling time-series data, highlighting the importance of effective windowing techniques. Various fixed window sizes (10, 20, 25, 30, 40, 50-time cycles) with consistent overlap were tested to isolate the window size as the primary variable. Results indicate that larger window sizes generally improve the performance of LSTM models by capturing essential temporal dependencies, while CNN models perform optimally with smaller windows. The paper underscores the trade-offs between window size, accuracy, and computational efficiency, offering insights into optimal configurations for reliable RUL predictions in predictive maintenance applications.