Semi-supervised adversarial discriminative learning approach for intelligent fault diagnosis of wind turbine

Abstract

Wind turbines play a crucial role in renewable energy generation systems and are frequently exposed to challenging operational environments. Monitoring and diagnosing potential faults during their operation is essential for improving reliability and reducing maintenance costs. Supervised learning using data-driven techniques, particularly deep learning, offers a viable approach for developing fault diagnosis models. However, a significant challenge in practical wind power equipment lies in the scarcity of annotated samples required to train these models effectively. This paper proposes a semi-supervised fault diagnosis approach specifically designed for wind turbines, aiming to address this challenge. Initially, a semi-supervised deep neural network is constructed using adversarial learning, where a limited set of annotated samples is used in conjunction with a vast amount of unannotated samples. The health status features present in the unannotated samples are leveraged to capture a generalized representation of the underlying features. Subsequently, a metric learning-guided discriminative features enhancement technique is employed to improve the separability of different manifolds, thereby enhancing the performance of the semi-supervised training process. By employing this methodology, it becomes possible to develop a fault diagnosis model with superior accuracy using only a limited amount of annotated samples. Comprehensive fault diagnosis experiments were conducted on a wind turbine fault dataset, revealing the efficacy and superiority of the presented methodology.