Contribution Imbalance and the Improvement Method in Multisensor Information Fusion-Based Intelligent Fault Diagnosis of Rotating Machinery

Abstract

The contribution of different signals to rotating machinery fault diagnosis can vary significantly, leading to suboptimal performance in multisensor information fusion-based intelligent fault diagnosis (MIF-IFD). This article examines the issue of imbalanced contributions in MIF-IFD models, explores its causes, and proposes an improvement method. We introduce a contribution discrepancy module to evaluate the contribution of various sensor signals to fault identification. By controlling the training pace of high-contribution branch networks, low-contribution parts are trained sufficiently to keep up. In addition, a distillation module is added to guide each branch network’s learning direction by using outputs from pretrained single-sensor networks as supervisory signals. This approach helps mitigate the degradation in feature extraction ability due to imbalanced training. Experimental results show that the proposed method performs well across two datasets and is valuable for practical deployment in MIF-IFD systems.