Optimum feature selection for the supervised damage classification of an operating wind turbine blade

There exists several methods for feature extraction from monitored infrastructure data. However, for wind turbine monitoring, there exists a dearth in the analysis of full-scale tests for damage detection, guidance for feature extraction, and expected performance levels. This paper attempts to address this through a benchmark study of feature selection and reduction methods, and their performance levels using a full-scale experimental study. Optimum feature selection is achieved through vibration-based damage detection of wind turbine composite blades using acceleration measurements. Experimental measurements of accelerations at 11 different locations on an operating Vestas V27 wind turbine is used. Variable operating conditions for three different rotor speeds corresponding to one healthy, three damaged, and one repaired condition were analyzed. Feature selection and related benchmarking identified the minimal feature set with the best capability to distinguish between multiple damage states. Time, frequency, and time–frequency features were extracted from the accelerations for feature estimation. Feature selection performance was tested via seven machine learning approaches using impulse and ambient responses. Results demonstrate how high classification accuracies of up to 95% can be achieved with only top 10 time–frequency features selected from the analysis of postimpact impulse responses. However, a larger set of optimal features between 30 and 50 are required for effective damage detection with ambient responses. The work, for the first time, presents such feature selection and classification performance for real-time wind turbine blade damage and as a function of optimum selected features. It also demonstrates how a reduced set of features is useful in distinguishing between damage states while a larger feature set may not be useful or even lead to reduced accuracy. The work provides a much needed performance benchmark for structural health monitoring practitioners and researchers for choice, comparison, validation, and use of features for wind turbines.