Robust Feature Learning for Remaining Useful Life Estimation Using Siamese Neural Networks

Abstract

Remaining useful life (RUL) estimation is very critical for planning the maintenance of machinery in various industries. Deep learning models have gained popularity as the key tools in estimating RUL. However, raw sensor measurements are affected by multiple factors beyond the degradation level, which is the primary goal of RUL estimation. This adversely affects model training, especially in the lack of sufficient data. To address the aforementioned issues, we propose to train an artificial neural network (ANN) that captures the smooth and monotonous degradation process. A siamese neural network (SNN) model is used to train the ANN so as to minimize feature variation across consecutive time instances while maximizing it across changes in health conditions. The effectiveness of the smooth features in RUL estimation is illustrated using turbofan engine degradation simulation data set and degradation image stream data set that are collected from rotating engines. We further discuss that the new features can be individually used to assess the health condition of the machinery without an RUL estimator.