Iterative fine-grained decoupling graph neural network for offshore wind turbine remaining useful life prediction

The effectiveness of graph neural networks (GNNs) in handling spatial dependencies among sensors has driven their adoption for remaining useful life (RUL) estimation. Yet, present GNN-based approaches directly apply node features to RUL prediction and sensor spatial dependency modeling, neglecting the distinct requirements of these two tasks for node features and failing to achieve task-level feature optimization. Therefore, this paper proposes an iterative fine-grained decoupling graph neural network (IFGDGNN) with adaptive graph pooling for multi-sensor equipment RUL prediction. In IFGDGNN, the proposed node feature iterative decoupling module (NFIDM) employs an iterative pre-decoupling strategy and a decoupling enhancement function, achieving fine-grained decoupling of node features. Subsequently, the decoupled features-based graph update module (DFGUM) is proposed to use the decoupled features separately for spatial dependency modeling and RUL prediction. DFGUM provides task-specific feature optimization directions for NFIDM and employs a coupling network to enable iterative graph updates. Furthermore, a node connectivity-based adaptive graph pooling module is constructed to discard weakly correlated nodes adaptively, optimizing the graph's readout process. The proposed method has not only achieved outstanding performance on two public datasets, but has also been successfully applied to the RUL prediction of the real-world offshore wind turbine, thereby demonstrating its engineering value. • A new iterative fine-grained decoupling graph neural network is put forward. • A node feature iterative decoupling module is designed to decouple node features. • A decoupled features-based graph update module is proposed for graph update. • An adaptive graph pooling module is proposed to reduce the redundant information.