A novel spiking graph convolutional residual network for robust fault diagnosis of planetary gearboxes under strong noises

Abstract

Fault diagnosis of planetary gearboxes is crucial for ensuring the safe and stable operation of key equipment such as mining machinery and wind turbines. However, in actual industrial environments, strong background noise often overwhelms weak fault features, leading to a significant performance degradation of traditional deep learning-based diagnostic methods. To address this challenge, an intelligent diagnostic method based on a spiking graph convolutional residual network is proposed, effectively achieving deep fusion and enhanced discriminability of the spatio-temporal features of fault signals. First, the integrate-and-fire neurons are embedded into the graph convolutional layer to construct a dynamic spatio-temporal fusion spiking graph convolutional layer to realize the synchronous extraction and spike coding of spatio-temporal features in the signal. Second, the channel attention enhanced spiking graph convolutional residual block is designed to capture the temporal modulation features by simulating the short-term memory mechanism of biological neurons, and the channel attention mechanism is used to enhance the expression ability of weak fault features. Finally, fault classification is achieved through a fully connected layer based on leaky integrate-and-fire neurons, which enable temporal accumulation of spikes. Experimental results demonstrate that the proposed method exhibits strong adaptability under conditions of sudden load, different types of signals, data imbalance, and strong noise environments. Notably, it maintains a diagnostic accuracy of 94.2%–96.9% even under intense noise conditions of −10 dB, significantly outperforming traditional diagnostic methods. This study provides a highly robust solution for fault diagnosis for planetary gearbox fault diagnosis in the presence of strong noise disturbances in industrial sites.