Dual-branch spatio-temporal graph network for bearing fault diagnosis

Bearings are one of the critical components in rotating machinery. Bearing failures can lead to equipment damage, reduced performance, and even major safety accidents. Therefore, improving the ability to diagnose bearing faults can help improve the availability, reliability, and safety of rotating machinery. However, the original vibration signals in rotating machinery often contain noise and irregularities, making it difficult for traditional vibration analysis to extract effective high-dimensional features. Inspired by the construction of spatio-temporal graphs and dual-branch graph networks, a bearing fault diagnosis method based on dual-branch spatio-temporal graph networks (DBSGN) is proposed. Firstly, the vibration signal is modeled based on spectrum theory and the spectrum analysis method to construct a spatio-temporal graph. Secondly, Laplace-based spectral decomposition is used to extract the feature vectors of samples in the spatio-temporal graph. Finally, we designed a dual-branch fusion network to train and verify the bearing data and adjusted the model's learning of the bearing data through a dynamic attention mechanism. The experimental results on three benchmark datasets indicate that DBSGN outperforms traditional models in terms of stability and accuracy.