A Domain Adaptive Convolutional LSTM Model for Prognostic Remaining Useful Life Estimation under Variant Conditions

In the age of industry 4.0 and smart manufacturing, a large volume of sensor data is produced from cyber-physical systems (CPS) and prediction of remaining useful life (RUL) of a machine or system becomes crucial for prognostics and health management (PHM). Several linear regression and deep learning models have been studied to extract features from segmented time windows and learn the degradation patterns. However, distributions of features are varying between source learning domain and target test domains, due to different working conditions and environments. Thus, the generalization of traditional methods will be influenced, which leads to performance degradation. This paper develops a domain adaptive CNN-LSTM (DACL) model to predict the RUL of a system based on the multi-dimensional sensor data. The DACL model combines the CNN and LSTM with domain adaptive transfer mechanism and take the operating conditions into consideration. The features extracted by CNN of both source and target data are transformed to a higher dimensional space by reproducing kernel Hilbert space (RKHS) and the loss function is compensated by using maximum mean discrepancy (MMD) to reduce the distributions discrepancy. The model is evaluated on C-MAPSS dataset and demonstrate its performance improvement by comparing with previous methods.