Remaining Useful Life Prediction under Multiple Operation Conditions Based on Domain Adaptive Sparse Auto-Encoder

Abstract

In the industrial production process, the remaining useful life (RUL) of the machine part is the key factor to determine the product quality, so it is important to predict the RUL of the machine part for industrial manufacturing. With the development of intelligent manufacturing, data-driven RUL prediction has become very popular. When the training dataset and the test dataset are distributed similarly, the traditional machine learning prediction method is very effective. However, in actual production, the operation conditions of the machine part used for training and testing may be different, resulting in different distribution of data sets. In this paper, we propose a domain adaptive SAE-LSTM (DASL) model for RUL prediction of the machine part to solve this problem. The DASL model contains sparse autoencoder (SAE) and Long Short-Term Memory (LSTM) with domain adaptive mechanism. The latent features extracted by SAE from source dataset and target dataset are transformed to reproducing kernel Hilbert space (RKHS) and the distribution discrepancy is reduced by using maximum mean discrepancy (MMD). Then the latent features are input into the LSTM to predict the RUL. What is more, the case where both source data and target data are data containing multiple conditions is also considered. The proposed model is tested on Foxconn tool wear dataset and PHM Challenging 2012 dataset. The results show that the method has a better improvement. In most experiments, this method outperforms other state-of-arts' methods.