A Deep Domain Adaptative Network for Remaining Useful Life Prediction of Machines Under Different Working Conditions and Fault Modes

Abstract

As one of the key techniques in prognostic health management (PHM), remaining useful life (RUL) prediction of machine relies on sufficient prior observed degradation data. Most previous RUL prediction methods assume that feature distributions of both datasets from source and target domains are similar. However, the data distributions are different across different domains because of different operating conditions and fault modes, which raises the performance deterioration of these methods in machine RUL prediction. In this article, a novel transfer learning method, deep domain adaptative network (DDAN) is proposed for machine RUL prediction that can handle feature distribution shift across domains under different working conditions and fault modes. First, a novel feature extractor, selective convolutional recurrent neural network (SCRNN), is developed for feature extraction from sensor signals. The selective convolutional recurrent blocks in SCRNN are able to effectively extract temporal and spatial features from vibration signals simultaneously. It incorporates heterogeneous temporal data of multiple sensors in the RUL prediction task. A collaborative domain alignment method is proposed to guide SCRNN to learn domain-invariant features of source and target domains. The DDAN-based RUL prediction method is evaluated on the systematic dataset, i.e., NASA Commercial Modular Aero-Propulsion System Simulation (C-MAPPS), and component dataset, i.e., PHM Challenging 2012 bearing dataset. The testing results illustrate the effectiveness of DDAN in machine RUL prediction under different operation conditions.