A digital twin-based approach for generalized fault diagnosis of high-speed train gearbox gears

Data-driven methods have been widely applied to fault diagnosis of high-speed train gearbox gears; however, their performance heavily depends on high-quality training data. Due to the inevitable degradation of gears during long-term service, significant distribution discrepancies arise between real-time operational data and historical data, which degrade the generalization capability and diagnostic accuracy of models. To address this issue, this paper proposes a digital twin-based generalized fault diagnosis method for high-speed train gearbox gears. First, a gearbox dynamic model is established using multibody dynamics and Hertzian contact theory,capable of characterizing operational states and generating vibration data under different health conditions. Second, an interactive generative adversarial network–based data fusion framework is developed. Through an interactive training strategy, the distribution gap between data in the digital and physical spaces is reduced, enabling collaborative learning of fault mechanism information and environmental characteristics. Finally, a deep convolutional neural network is trained using the fused data to achieve high-accuracy generalized fault diagnosis. Experimental results demonstrate that the proposed method achieves an average diagnostic accuracy of 96.70%, outperforming the best comparative method by 2.20% and conventional methods by 4.67%, thereby validating its effectiveness and superiority in improving generalization performance for fault diagnosis.