The data-driven multi-sensor graph neural network (GNN) method is one of the key technologies for rotating machinery fault diagnosis. This paper innovatively proposes a selective and cooperative serial expansion GNN method for fault diagnosis. It aims to address two key challenges in multi-sensor scenarios: the inherent limitations of the message-passing mechanism in the existing GNN-based fault diagnosis methods, and the dynamic node variations caused by sensor malfunction. For the periodic characteristics presented by the vibration time-series signal of the sensor nodes, this study proposes a series expansion graph network architecture based on Fourier theory, which can effectively enhance the model’s capability in capturing the periodic fault features. Meanwhile, a selective and cooperative strategy is proposed to address the key issues of static data fusion in GNN and the dynamic change of the effective number of sensors. This strategy utilizes a behavior network to determine each sensor node’s behavior status during message-passing and introduces a node selection network to evaluate sensor reliability. The message-passing network realizes dynamic cooperative feature fusion based on the nodes’ behavioral status and reliability information obtained above. Finally, a selective and cooperative Fourier series expansion graph fault diagnosis method with excellent adaptivity and robustness is proposed. The two experimental cases show that the proposed method has a diagnostic accuracy rate of 99%, even under variable rotational speed conditions. More importantly, it can achieve self-adaptation in the event of sensor node failure and maintain a diagnostic accuracy rate of at least 90%.
He et al. (Mon,) studied this question.