Multi-similarity and gradient fusion digital twins for fault detection and diagnosis of rolling bearings

Abstract

In rolling bearing application scenarios, the challenges of acquiring faulty data have led to research focusing on unsupervised fault detection and diagnosis methods trained solely on healthy data. In this study, we built a deep digital twin of a healthy rolling bearing state by combining multi-similarity metrics and a model backpropagation mechanism to fully mine fault information and achieve advancements in both fault detection and diagnosis. Our proposed approach establishes a fault scoring metric set (FSMS) by integrating multi-similarity metrics and model gradient information. Furthermore, a selection and fusion strategy for the FSMS is designed based on the old stage generated validation datasets to obtain a fusion fault scoring metric and realize fault detection. A gradient fusion digital twin is further proposed for fault diagnosis. The method fuses time–frequency and model gradient features to distinguish different fault modes. To verify the effectiveness of the proposed method, experiments were conducted on rolling bearing datasets. The experimental results show that the proposed method has excellent performance, effectively integrating the fault information embedded in multi-similarity metrics and gradient information, while exhibiting strong robustness and generalization to variations in hyperparameters. This study provides a promising new idea for digital twins in fault prediction and health management of rolling bearings.