An Ensemble Data-Model-Label Three-Level Regularization Framework for Imbalanced Intelligent Fault Diagnosis

Abstract

In real industrial scenarios, fault data are characterized by class imbalance, a major challenge for data-driven intelligent fault diagnosis. This article proposes a novel three-level regularization framework that integrates data, models, and labels to diagnose the imbalanced fault. First, a signal to image (S2I) module is introduced, which converts 1-D signals into 2-D images to conduct research and reduce model development workload and model-specific dependencies. Then, a regularization framework is proposed consisting of three submodules, inner local feature regularization (ILFR), outer local feature regularization (OLFR), and class balance margin loss (CBML), improving the faulty health state recognition accuracy without degrading the normal health state recognition performance. Finally, adequate experiments are carried out on four mechanical fault datasets. The results show that under the extremely imbalanced conditions, the proposed framework can improve the accuracy of the baseline method by 28%, 38%, and 25% on the three datasets (including PU, JNU, and UoC), respectively. Moreover, the proposed framework outperforms the SOTA method on the CWRU dataset, which validates the effectiveness and superiority of the proposed framework.