Remaining Useful Life Prediction Using Deep Multi-scale Convolution Neural Networks

Abstract

Abstract Accurate Remaining useful life (RUL) prediction is the premise of system prognostics and health management (PHM). In fact, it is often difficult to predict the accurate RUL because of the diversity of working condition and self-state between systems. In this paper, a data-driven method is proposed for RUL prediction using deep multi-scale convolution neural networks (DMSCNN), which is made up of four convolution layers. Two of the four convolution layers are multi-scale convolution layers, which are composed by three different scales convolutions connected by deep concatenation layer. After the four convolution layers, we add fully connected layer and regression layer to build a deep learning structure for predicting RUL. The IEEE PHM 2012 data challenge bearing dataset is employed to verify the proposed method. The experiments result show that the proposed DMSCNN has better performance and higher prediction accuracy than other methods. Additionally, the t-distributed stochastic neighbour embedding (t-SNE) dimension reduction method is used to reduce the computation. Experiments result demonstrate that the features processed by t-SNE have better distinguishability, and the prediction accuracy can be further improved by inputting the t-SNE features into the DMSCNN proposed in this paper for RUL prediction. This paper offers a new and effective approach for RUL prediction.