Digital twin-driven imbalanced fault diagnosis method based on new distribution discrepancy metric and large margin aware focal loss

Abstract

In actual industrial production, the frequency of gear failures is much lower than that in normal conditions. The scarcity of fault samples leads to a severe data imbalance problem, which significantly limits the performance of deep learning-based fault diagnosis methods. To address this issue, this paper proposes a digital twin-driven imbalanced fault diagnosis method based on a New Distribution Discrepancy Metric (NDDM) and Large Margin aware Focal (LMF) loss. First, a fault virtual data generation strategy based on digital twin technology is proposed. By analyzing the nonlinear dynamic characteristics of the gearbox, an effective virtual model of the gearbox is established, generating a large amount of high-quality virtual fault data to mitigate the data imbalance problem. Then, the NDDM is employed to simultaneously align the marginal distribution and subdomain conditional distribution by reducing the distribution discrepancy between the virtual and actual domains. Finally, the LMF is adopted to further enhance the model's fault diagnosis performance by focusing on hard samples and preserving more inclusive decision boundaries for fault categories. Experimental validation on two datasets demonstrates that the proposed method significantly outperforms other approaches in handling imbalanced data, providing a novel solution for effective gear fault diagnosis under data imbalance.