A Two-Stage Data-Driven Method to Identify Blade Faults for Wind Turbines Using Vibration Data

With wind generation projected to constitute up to an estimated 12.1% of global power capacity by 2024, the industry grapples with the economic and safety challenges posed by blade faults in wind turbines. Vibration detection is one of the solutions towards the detection of blade faults. However, based on single-source vibration signals, it lacks an understanding of fault characteristics and scalable identification solutions. This paper presents a novel two-stage methodology for the detection of blade faults using single-sensor vibration data. The proposed method first classifies vibration signals into healthy or faulty states using central moment values as probability features, followed by a machine learning model to predict the blade’s general state. In that, this paper also uncovers the optimal data volume necessary for over 90% accurate state identification, potentially informing industry standards and reducing computational demands. The bottom stage develops an extreme value division method. Through training, this method could extract a threshold that maximizes the differences of vibration characteristics among these categories, thus differentiating the characteristics between four types of blade faults. These characteristics are visualised as the comparative set by fitting extreme value distributions. Field engineers can utilize the comparative set to fast identify fault categories for new vibration data without a given state.