Fatigue life prediction of variable-load industrial robotic arms based on physics-informed graph neural networks

Introduction Industrial robotic arms operate under long-term variable-load conditions, and joint fatigue accumulation may lead to sudden failures. Conventional life-prediction methods cannot effectively integrate spatial topology, temporal degradation rules and fatigue physical mechanics. This study proposes a physics-informed Graph Attention-GRU (PI-GAT-GRU) model for fatigue-life prediction of variable-load industrial robotic arms. Methods Robotic-arm joints are mapped into a topology-mechanics graph. Spatiotemporal features are extracted via dynamic graph attention and GRU units. The Miner damage criterion and stiffness-degradation model are embedded, and a physics-regularized joint loss function is constructed for end-to-end fatigue-life prediction. Comparative experiments against three baseline models and ablation studies are conducted on multi-body simulation datasets covering low, medium and heavy-load conditions. Results The subset-level prediction accuracy of PI-GAT-GRU reaches 96%–97% under low-load, 94%–97% under medium-load, and 93%–95% under heavy-load conditions. Its overall average accuracy on the mixed full-condition test set is 92.48%. Benchmark models achieve 91%–94% accuracy under low-load and drop to 88%–92% under heavy-load conditions. The proposed model obtains better accuracy and robustness under variable-load scenarios. Discussion The PI-GAT-GRU realizes effective fusion of data-driven learning and fatigue physical constraints. It supplies reliable technical support for predictive maintenance and health management of industrial robotic arms. Further work will focus on cross-platform validation and nonlinear-damage mechanism integration.