Selecting causal features with higher-order suppression mechanism in compound bearing faults: a perspective of hyper-networks

Abstract In recent years, complex network analysis has gained widespread application in fault diagnosis and prediction for rolling bearings, particularly in compound fault feature selection. However, the features of compound fault signals are interrelated across different dimensions due to nonlinear superposition, which often results in masking weak fault signal information when hidden interactions among features occur. To address this, we present a novel PageRank algorithm for compound fault feature selection by integrating a higher-order suppression mechanism with hyper-network theory. In this method, the causal relationships among compound fault features are first represented by a hyper-network model. Subsequently, the PageRank algorithm with a higher-order mechanism is established by using information from hyperedges of different dimensions. The criteria of feature selection are provided by the accuracy of fault diagnosis through deep neural network models. The effectiveness of the proposed method is empirically validated through experiments on synthetic and real-world compound fault datasets, demonstrating its capability in fault diagnosis and prediction. The experimental results demonstrate that the compound fault feature selection method with higher-order PageRank outperforms traditional approaches, showing significant improvements in both diagnostic accuracy and predictive capability for compound faults in rolling bearings.