Rotating Machinery Prognostics via the Fusion of Particle Filter and Deep Learning

Abstract

Prognostics and health management (PHM) emerges to be a promising technology which enhances the reliability and reduces maintenance cost of rotating machineries. Prognostics, as the key part of Structural Health Monitoring (SHM), focuses on predicting the remaining useful life (RUL) of a system, based on the current and historical monitoring data. RUL prediction algorithms can be roughly divided into two categories: model-based method and data-driven methods. Among the two, data-driven methods are widely used because of the availability of monitoring data and difficulty of developing accurate mathematic physical models. However, traditional data-driven methods ignore the physical mechanism and have limitation in uncertainty modelling of predicted results. To overcome the weakness of data-driven methods and improve the accuracy of prediction, a novel fusion algorithm of particle filter (PF) and long shortterm memory (LSTM) neural network is proposed in this work. By this means, the RUL prediction from LSTM is obtained with confidence intervals through PF. The proposed method is validated using crack size propagation and the RUL prediction using traditional LSTM is presented for comparison. The results demonstrate the effectiveness of the fusion algorithm which integrates the advantages of both PF and LSTM.