Fault prediction method towards gearbox based on digital twin and deep transfer learning

Abstract Due to the inherent limitations of measured data and the strong data dependency of deep learning models, there is growing attention on using Digital Twin technology to generate simulated data for model training. A critical technical challenge lies in ensuring the consistency between virtual and real-world data—a key prerequisite for advancing Digital Twin-driven gearbox health monitoring and intelligent operation and maintenance. Thus, developing novel intelligent fault diagnosis methods for gearboxes holds considerable significance. Traditional data-driven fault diagnosis approaches generally assume identical distributions between training and testing datasets and require substantial amounts of historical data to build reliable diagnostic models. However, such assumptions are often unrealistic in the dynamic and variable operational environments of real-world industrial gearbox systems. The emergence of Digital Twin and transfer learning technologies presents a promising pathway toward intelligent manufacturing, offering the potential to overcome the constraints of existing methods and establish a new paradigm for gearbox fault diagnosis. This study integrates model-based knowledge with data-driven techniques by leveraging Digital Twin and transfer learning to perform gearbox fault diagnosis. First, a dynamic Digital Twin model of the gearbox is constructed using MATLAB/Simulink to simulate fault data under diverse operating conditions—data that are difficult to obtain from actual operations. Subsequently, a multi-level convolutional neural network (MLCNN)-based algorithm is proposed to reduce the cumulative errors often introduced during conventional data preprocessing. The model is then pre-trained using a large, balanced dataset generated via the Digital Twin. Finally, transfer learning enables the application of the trained model to real-time fault diagnosis in industrial robot gearboxes, achieving a transition from virtual simulation to physical deployment. Experimental results confirm the feasibility of the proposed approach, demonstrating a diagnostic accuracy of 97.85%—an improvement of 14.3% over conventional convolutional neural network methods without Digital Twin integration.