Power Spectral Entropy-based Graph Construction for Rotating Machinery Diagnosis Using Multi-sensor

Multi-sensor fusion-based methods have achieved great results in rotating machinery fault diagnosis. However, many of them fail in mining the relationship between signals during the fusion process, leading to the loss of information. Graph convolutional network (GCN) can effectively mine the relationship between nodes hidden in the geometric structure, which can further improve the performance of multi-sensor fusion-based methods. Therefore, a power spectral entropy-based graph construction for rotating machinery diagnosis using multi-sensor is proposed. Power spectral entropy is extracted from multiple sensors using wavelet packet decomposition. Then, a fused feature is constructed by combing them. On the basis, the power spectral entropy-based graph is obtained by a distance-based clustering algorithm using the fused feature. Finally, graph feature learning with GCN is used to extract fault features. Verification experiments were conducted on a gearbox experimental platform, and results show that the proposed method can effectively classify faults.