Remaining useful life prediction based on graph feature attention networks with missing multi-sensor features

Abstract

Prognostics and health management (PHM) is important to ensure the reliable operation of industrial equipment, where monitoring the degradation process of machinery through multi-source sensors for remaining useful life (RUL) prediction is one of the key tasks. In recent years, deep learning-based time series forecasting methods have been proposed to predict RUL as they have strong capability on temporal correlation modeling for time series gathered by sensors. However, these methods usually operate under the assumption of a static, fixed-dimensional feature set. The proliferation of sensors inevitably escalates the probability of missing and anomalous features within measurement data, thereby causing the dimensions of input features to dynamically fluctuate over time. Therefore, this paper proposes a Graph Feature-Gated Graph Attention Network (GF-GGAT), which is capable of fusing multi-sensor data with partially missing sensor data and performing RUL prediction. First, the problem of spatio-temporal map construction when some sensor data are missing is solved by introducing dynamic time regularization. Second, the feature-deficient multi-sensor data are inductively learned through graph feature transformation and stepwise graph convolution. Finally, spatio-temporal features are extracted by a gated graph attention network (GGAT) to accomplish RUL prediction. Two case studies demonstrate the superiority of the proposed method over state-of-the-art RUL prediction methods.