Unsupervised Prognostics based on Deep Virtual Health Index Prediction

Abstract

Prediction of the Remaining Useful Life (RUL) for industrial systems has been facilitated by the acquisition of large amounts of real-time data and the use of deep learning methods. However, the vast majority of these methods rely on the availability of extensive RUL-labeled data, which is not the case for most of real industrial applications. The goal of this paper is to show how unsupervised learning can provide alternative ways to address this issue. The proposed method is essentially made of two steps. First, a Virtual Health Index (VHI) is extracted in an unsupervised manner from the raw sensor data using a Deep Convolutional Neural Network (CNN) autoencoder. Secondly, an Long-Short Term Memory (LSTM) Encoder-Decoder predicts the future values of the VHI, until an End-of-Life (EOL) pattern is recognized (using a sliding window DTW algorithm). The suggested method is tested on the C-MAPSS dataset and offers promising results with a great potential to be applicable on real-life use cases.