Open Access

Yi QinORCiD, Linjie ZhengORCiD, Jun Luo, Junyu QiORCiD, Yongfang MaoORCiD

• Graph neural networks (GNNs) have been widely applied to remaining useful life (RUL) prediction driven by multi-source information fusion. However, the conventional GNN-based approaches use the shared node features for both RUL prediction and spatial dependency modeling, which may induce implicit information conflicts. Moreover, they have no effective mechanisms for identifying the weakly correlated nodes, which causes the prediction accuracy to be easily disturbed by noise. To resolve potential information conflicts, we propose a novel node feature contrastive decoupling module (NFCDM). The NFCDM employs a cross-fusion-contrastive decoupling strategy to explicitly decompose node features into global degradation features for RUL prediction and local relational features for spatial dependency modeling. Task-specific contrastive decoupling constraints are further designed to ensure the task relevance and information independence of the decoupled features. In addition, to achieve the effective node screening, a deviation-diffusivity cooperative node selector (DDCS) is proposed. This novel selector is capable of constructing joint evaluation metrics based on the deviation of global degradation features and the information diffusion capability of nodes to adaptively discard redundant nodes. Built upon these two novel modules, the contrastive decoupling graph neural network with deviation-diffusivity cooperative node selector (CDGNN-DDCS) is proposed for RUL prediction. Comparative experiments on public datasets confirm the superiority of the proposed method. Lastly, its engineering practicality is successfully verified via an RUL prediction experiment on actual wind turbines.