Digital twin-driven focal modulation-based convolutional network for intelligent fault diagnosis

Rolling bearings are essential components of various rotating machinery and are critical in ensuring safe and reliable industrial production. Deep learning techniques have demonstrated outstanding potential for real-time monitoring of bearings, contributing to the safe operation of machinery and equipment. However, deep learning-based fault diagnosis methods typically rely on training datasets comprising samples of all potential failure modes that may not be acquirable in specific industrial settings. To tackle the challenge above, this paper introduces a digital twin approach to generate synthetic data to supplement and enhance the quality and availability of training data in deep learning methods. Specifically, the main contributions of this research are: (1) constructing a digital twin model of rolling bearings to generate an approximation of the physical entity bearing status data. (2) investigating the efficient combination of CNNs and focal modulation mechanism, and proposing a novel lightweight architecture, FM-LCN, aims to learn local-global representations of simulated data to improve diagnostic performance. Experiments demonstrate that FM-LCN outperforms five state-of-the-art competitive models by a large margin in accuracy with lower computational cost.