DCAGGCN: A novel method for remaining useful life prediction of bearings

Abstract

Accurate prediction of Bearings' remaining useful life (RUL) is crucial in equipment operation and maintenance. The bearing RUL prediction technology based on GCN has recently been widely used. However, the existing GCN-based RUL prediction results are limited by two aspects : (1) GCN usually uses the predefined adjacency matrix to define the graph, which makes the graph unable to track the real-time correlation of degradation features in time. (2) Existing GCN uses only one to two layers of graph convolution and cannot extract deep features. Based on the issues above, this paper proposes a bearing RUL prediction model that utilizes a Dual-correlation adaptive gated graph convolutional network (DCAGGCN). Firstly, a predefined double correlation graph is proposed and obtained by feature channel data. Next, an adaptive graph is created by transforming a source matrix and a target matrix, and then integrating it with a predefined graph. This allows the network to consider two types of correlation and adaptively adjust the graph's topology. In addition, this paper proposes a gated convolution layer, which can greatly alleviate the over-smoothing problem caused by the stacking of graph convolution layers. The effectiveness of the proposed method is verified by two public datasets.