An improved similarity-based prognostic algorithm for RUL estimation using an RNN autoencoder scheme

Abstract

Remaining useful life (RUL) estimation of a degrading system is the major prognostic activity in many industry applications. This paper presents an improved version of the similarity-based curve matching method for the remaining useful life estimation of a mechanical system, which is a companion paper of our previous work on RUL estimations using a bidirectional recurrent neural network (RNN) based autoencoder scheme. We propose a zero-centering rule to tackle the varying initial health across instances (systems) when using the similarity-based health index curve matching technique to identify the training instances that share a similar degradation pattern with the test instance whose RUL needs to be determined. However, this rule will also induce a significant prediction error, especially when the off-line training instances are abundant, or the true RULs of the on-line test instances are large. Thus, an ensemble approach that integrates the RUL estimations obtained from the similarity-based curve matching techniques, with and without the zero-centering rules, is introduced to increase the robustness and accuracy of proposed method for RUL estimations. We evaluate the prognostic performance of the ensemble algorithm and standalone algorithms on four publicly available turbofan engine degradation datasets. The results demonstrate that the proposed ensemble approach gives more robust and reliable RUL estimations compared to any independent algorithm used on all the studied datasets.