Neural Networks for Predicting Fluid Filter Remaining Useful Life

Abstract

This research addresses the challenge of estimating the Remaining Useful Life (RUL) of oil filters in industrial hydraulic systems using data-driven predictive maintenance. Focusing on a proprietary dataset characterized by a severely limited number of operational cycles and sparse laboratory measurements, the study evaluates traditional machine learning and deep neural networks under various feature engineering approaches. Findings reveal that for this constrained dataset, predictive accuracy is critically dependent on a single, dominant feature representing the filter’s total workload. Consequently, RUL defined by processed oil volume proved to be a more robust and predictable target than one based on operational time. While complex feature engineering and models struggled with the limited data, the same methodologies demonstrated strong performance on comprehensive benchmark datasets. To overcome data limitations in the target application, a market study for inline particle sensors was conducted, identifying feasible technologies that could provide the high-frequency oil cleanliness data necessary for robust future RUL predictions. The study underscores the fundamental importance of sufficient, relevant data for successful predictive maintenance implementation.