A Remaining Useful Life Approach using an Ensemble Regressor enhanced with Hilbert Transform and CNN-LSTM Model

Authors : Azeez Adebayo, Kaplan Kaplan, Hüseyin Metin Ertunç

Pages : 137-150

Doi:10.34088/kojose.1426751

View : 13 | Download : 12

Publication Date : 2024-11-30

Article Type : Research Paper

Abstract :In this study, we propose a hybrid approach that integrates signal-driven and knowledge-based techniques to estimate the Remaining Useful Life (RUL) of bearings. The experimental data for this research is sourced from the FEMTO-ST Institute. Firstly, the horizontal and vertical acceleration data is ordered chronologically by time, and a band-pass filter is used for early-stage preprocessing of the vibration signals below 20 kHz. Then, the overall behavior of the signal is characterized by Hilbert-Transform. For the feature extraction scheme, a model that integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks is implemented. These features form historical data on health indexes describing fault stages and are as such used to fit a voting regressor yielding an extrapolated future. The voting regressor is based on support vector regression (SVR) and linear regressor methods and a fault threshold is determined as 0.8 based on prior experiments. Finally, the proposed methodology distinguishes itself by recording the smallest average percentage error on the FEMTO dataset. This method proves that early-stage predictions are possible with run-to-failure data provision ranging from 60% and above, averaging some 1400 seconds into the future implying its suitability and effectiveness for real industrial applications.
Keywords : Predictive Maintenance, CNN LSTM, Feature Extraction, RUL, Hilbert Transform