A physics based data generation method using a dynamic model for wind turbine gearbox fault diagnosis in fault data scarce settings

Machine-learning-based condition monitoring of wind turbine gearboxes is critical for enhancing gearbox reliability and operational safety. However, existing deep learning methods often rely on large-scale real-world datasets, which are typically difficult to obtain in industrial settings. To address this data-scarcity challenge, this study proposes an intelligent fault diagnosis framework that integrates physics-based modeling with data-driven learning. We develop a nonlinear dynamic model of a wind turbine gearbox and perform feedback calibration using the Pearson correlation coefficient (PCC) to improve the agreement between simulated responses and measured signals. We then embed fault-mechanism-based mathematical formulations into the calibrated model to generate fault samples, thereby compensating for the limited availability of real fault data. The generated dataset is used to train deep learning models, which are subsequently validated on previously unseen real fault measurements. Experimental results show that the proposed method achieves a fault classification accuracy of 97%. Under a limited real-sample regime, augmenting the few-shot training set with simulated samples further increases the fault classification accuracy to 98%, demonstrating that the proposed framework can effectively mitigate the degradation caused by scarce wind turbine gearbox fault data, thereby supporting mechanism-consistent failure identification in practical field applications.