A Transferable Topology-Aware Graph Pooling Network for Remaining Useful Life Prediction Under Cross-Domain Conditions
Jiusi Zhang, Hao Luo, Juncheng Hu, Chao Cheng, Shimeng Wu, Pengfei Yan, Jilun Tian
Abstract
Remaining useful life (RUL) prediction is critical for ensuring the safety, reliability, and efficient operation of complex industrial systems. Conventional deep neural networks, such as long short-term memory network and convolutional neural network mainly adopt local recurrent and pooling operations, which have notable limitations in global modeling in RUL prediction for complex industrial systems. Although the emergence of graph neural network (GNN)-based approaches can solve the above-mentioned problems, there are few reports on using GNN for RUL prediction under cross-domain conditions. Consequently, a novel transferable topology-aware graph pooling network is proposed for RUL prediction under cross-domain conditions. Specifically, this article integrates local and global voting to construct a topology-aware graph pooling network. Furthermore, a loss function is constructed to describe domain differences, so that the RUL prediction under cross-domain conditions is accomplished. Considering most of the existing deep learning approaches are closed-box models, this article designs an in-depth interpretability analysis based on the feature distribution and integrated gradient for the proposed approach. The effectiveness of TTAPGN is demonstrated by a commercial aircraft turbofan engine dataset, and a bearing degradation dataset from Xi'an Jiaotong University.