A class-imbalance-aware domain adaptation framework for fault diagnosis of wind turbine drivetrains under different environmental conditions

This paper proposes a new fault diagnosis model for wind turbine drivetrains addressing time-varying environmental conditions and the limited availability of fault-related data. Wind turbines continuously work under variable conditions, making it difficult to generalize a baseline-trained machine learning model to these conditions with different domain distributions. On the other hand, limited labeled data are usually available for failure events in real wind turbines, leading to class-imbalanced training data for the machine learning model. Particularly, offshore wind turbines face increasing challenges from harsh environmental conditions and dynamic wave loading affecting data collection. This study proposes an improved domain adaptation model called ”class-imbalance-aware deep adversarial adaptive network (CIDAA)” for wind turbine drivetrain fault diagnosis. This model utilizes a class-imbalance-aware layer to learn different domain-invariant features while improving the discriminative structure of the class-imbalanced feature space. Therefore, the model can efficiently generalize from a labeled source domain to an unlabeled target domain, even in the case of class imbalance. The proposed model is verified by a bearing damage dataset obtained from a high-fidelity 5-MW reference drivetrain model mounted on a spar-type floating wind turbine under three environmental conditions. The results demonstrate that the CIDAA model can diminish the effect of time-varying environmental conditions and class-imbalance data, leading to higher fault classification accuracy than standard models.