Adaptive Piecewise Linear Function and Deep Learning for Remaining Useful Life Estimation

A precise estimation of the Remaining Useful Life (RUL) of a machine is a crucial aspect of predictive maintenance. Triggering potential failures beforehand relieves planning maintenance operations, reducing downtime, and improving the logistics around spare parts. A common theme that is present across research is the use of the piecewise linear RUL function which is used to label the ground truth RUL for model evaluation. However, these methods are limited by the lack of an explainable approach to piecewise threshold selection. To estimate the RUL of a machine component, deep learning models such as temporal convolutional network (TCN), long short-term memory (LSTM) and their hybrid variants are utilized, incorporating an adaptive selection of the upper threshold for improved accuracy. In this approach, the Kneedle algorithm is used to obtain the adaptive thresholds for piecewise linear RUL function. The proposed methodology is experimented using the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) dataset. When comparing results with some of the existing statistical, regression and deep learning based methods, this work demonstrates comparatively higher performance.