Node-level pooled graph neural network with adaptive for bearing remaining useful life prediction

Abstract Remaining useful life prediction for bearings based on graph convolutional network (GCN) is increasingly gaining widespread attention. GCN, with their excellent capability in extracting correlation information, demonstrates significant advantages in capturing degradation features from condition monitoring data. However, current GCN methods typically rely on time-domain feature-based graph construction and predefined graph structures for information aggregation, leading to low-quality graph feature information and difficulties in dynamically tracking changes in feature correlations during the bearing degradation process, thereby affecting prediction accuracy. Meanwhile, existing prediction models often employ direct stacking of graph nodes in the graph readout phase, which can easily result in the loss of local information and weaken the model’s ability to capture deep features. To enhance the robustness of feature representation against noise and capture evolving feature correlations, this paper utilizes short-time Fourier transform combined with bidirectional long short-term memory networks to derive robust spatiotemporal features. Building upon this, a novel adaptive graph mechanism is introduced that dynamically adjusts the topology of the graph during the model training process, enabling the model to better adapt to changes in bearing degradation features and thus improving the robustness and accuracy of predictions. Furthermore, this paper designs a node-level clustering pooling module that captures local information in the convolved graph and performs effective node pooling, preserving more comprehensive feature information. Experimental validation on public datasets demonstrates that the proposed method outperforms other state-of-the-art methods in prediction performance.