Data-Driven RUL Prediction

Abstract

In this chapter, the data-driven RUL prediction methods for mechanical systems are presented. Since the deep learning algorithm has shown remarkbale advantages on prognosis problems in the current literature, the neural network-based methods are focused on in this chapter. First, the deep separable convolutional neural network-based RUL prediction method is introduced, which establishes a direct mapping relationship between raw monitoring data and RUL by implementing separable convolution and constructing information refinement units. Next, the recurrent convolutional neural network-based RUL prediction method is illustrated. A network with temporal memory capability is constructed using recurrent connections and gating mechanisms. At last, we present a multi-scale convolutional attention network-based RUL prediction method. By the integration of multi-scale representation learning strategy, the degradation information of the mechanical system can be extracted in different time scales. Throughout this chapter, experiments on multiple run-to-failure datasets are carried out, which validate the effectiveness of the presented methods.