A two-stage framework for identifying and characterizing wind turbine noise data and its validation by listening tests

The reliable identification of acoustically dominant wind turbine noise in field measurements is essential for analysing source-specific noise characteristics and sound propagation under real atmospheric conditions. Long-term acoustic data sets typically contain mixtures of wind turbine noise and competing environmental sounds, which makes robust automated identification challenging. This paper presents a two-stage framework for identifying wind-turbine-noise-dominated periods and specific wind turbine noise components in large acoustic data sets. In the first stage, time periods dominated by wind turbine noise are identified by combining statistical preselection criteria, turbine operating data, and a physics-based signal analysis that relates detected modulation frequencies to blade passing harmonics. In the second stage, the selected periods are further examined to identify specific wind turbine noise components, including rotor-induced amplitude modulation, tonal components, and high-frequency whistling noise. The framework is validated against a perceptual reference derived from a structured listening test in which wind turbine noise components and relevant competing noise sources are classified into predefined categories. A total of 15 participants evaluated audio segments, resulting in a reference with 166 classified minutes. The listening test shows good intrarater reliability (mean Jaccard index: 0.87) and moderate, category-dependent interrater agreement (mean: 0.56). Validation of the first stage demonstrates high performance for identifying dominant wind turbine noise (precision: 0.99, recall: 0.96). Component-specific validation of the second stage shows physically plausible detection behaviour, with deviations primarily attributable to subjective perception and masking effects in the listening test. The validated framework enables reliable and effective identification of wind turbine noise and its components, as demonstrated by its application to a 1-month acoustic data set containing interfering environmental noise.