A Graph Neural Network-Based Method for Predicting Remaining Useful Life of Rotating Machinery

Predicting remaining useful life of rotating machineries like gears / bearings accurately is vital to guarantee safe and reliable operation of equipments. With the development of sensor technology, more and more operation state signals of equipments could be collected effectively, thus enabling to achieve considerable development in data-driven prediction method of remaining useful life. Nevertheless, existing models only considered time sequences of samples, but ignores spatial information among sensors when processing health state degeneration data collected by multiple sensors. To address this problem, a deep adaptive spatial-temporal graph network model was proposed to predict remaining useful life of rotating machinery. Specifically, multiple state inspection information was preprocessed firstly through time window and each slice of each time window was divided into a remaining useful life value corresponding to one sample. Secondly, the model is divided into temporal convolution layer and graph convolution layer. The former one is composed of extended causal convolution and it is used to learn time sequence information. The later one contains the learnable adjacent matrix and it was used to learn spatial information of different-state detection data. After undergoing testing on a publicly available dataset, the model's evaluation metrics were found to be inferior to those of other high-performing prediction models. Moreover, validity of the graph convolution layer was verified through an ablation experiment.