Wasserstein Distance- EEMD Enhanced Multi-Head Graph Attention Network for Rolling Bearing Fault Diagnosis Under Different Working Conditions

Traditional fault diagnosis models often overlook the interconnections between segments of vibration data, resulting in the loss of critical feature information. Additionally, the vibration signals of rolling bearings exhibit non-linear behaviors during operation. Therefore, an efficient fault diagnosis model tailored for rolling bearings is proposed in this paper. In the proposed model, the 1D vibration signals are first preprocessed using ensemble empirical mode decomposition (EEMD). This technique generates multiple intrinsic mode functions (IMF) as individual nodes. The percentage distance between each node is calculated using the Wasserstein distance (WD) to capture the relationships between nodes and use it as the edge weights to construct a node graph. This unique approach enhances the transformation of 1D vibration signals into a node graph representation, preserving important information. An improved multi -head graph attention network (MGAT) model is established to extract features and perform classification on the node graph. This MGAT model effectively utilizes the relationships between nodes and enhances the accuracy of fault diagnosis. The experimental results demonstrate that the proposed method achieves higher accuracy compared to similar models while requiring less processing time. The proposed approach contributes significantly to the field of fault diagnosis for rolling bearings and provides a valuable tool for practical applications.