Remaining Useful Life Prediction Based on Multi-Representation Domain Adaptation

Abstract

All current deep learning-based prediction methods for remaining useful life (RUL) assume that training and testing data have similar distributions, but the existence of various operating conditions, failure modes, and noise lead to insufficient data with similar distributions during the training process, thereby reducing RUL prediction performance. Domain adaptation can effectively solve this problem by learning the cross-domain invariant features of the source domain and target domain to reduce the distribution difference. However, most domain adaptive methods extract the source domain and target domain features into a single space for feature alignment, which may leave out effective information and affect the accuracy of prediction. To address this problem, we propose a data-driven approach named long short-term memory network and multi-representation domain adaptation (LSTM-MRAN). We standardize and process the degraded sensor data with a sliding time window, use LSTM to extract features from the degraded data, and mine the time series information between the data. Then, we use multiple substructures in multi-representation domain adaptation to extract features of the source domain and target domain from different spaces and align features by minimizing conditional maximum mean difference (CMMD) loss functions. The effectiveness of the method is verified by the CMAPSS dataset. Compared with methods without domain adaptation and other transfer learning methods, the proposed method provides more reliable RUL prediction results under datasets with different operating conditions and failure modes.