Bearing fault diagnosis based on the TFS-D features and fusion neural network

Abstract Sensitive fault information is critical for bearing condition monitoring. In current research, time-domain and frequency-domain features are commonly used for diagnosis model construction. However, fault dynamics and deep cross-domain correlations are often neglected due to extraction difficulties. In this paper, a novel bearing fault diagnosis method based on time-frequency-spatial-domain (TFS-D) features and a fusion neural network (FNN) is proposed. The FNN is composed of two stages: data-level fusion and feature-level fusion. In the first stage, the fault dynamics of the rolling bearing are identified accurately via deterministic learning. In addition, multi-position dynamics is used to capture the coupling between vertical and horizontal vibration signals in the spatial domain, so that more sensitive information about internal system changes can be obtained. The extracted dynamic information, together with time-frequency fault data, is then fed into the FNN. In the feature-level fusion stage, the splicing direction and feature extraction angle are adjusted by a concatenation module, while local features extracted by CNN branches are fused with the global representation generated by the Transformer through a parallel concatenation module. Finally, the effectiveness of the proposed method is verified through offline and online bearing fault diagnosis experiments, and the superiority of the method based on comprehensive TFS-D information and the FNN is demonstrated.