Researchers trained software to pick out a wind turbine’s “fingerprint” from months of recordings filled with birdsong and wind
As wind turbines expand closer to homes and villages, scientists need to know exactly how turbine noise behaves under real atmospheric conditions — not in a lab, but outdoors, over months, in all kinds of weather. That requires long continuous recordings. The problem is that those recordings are cluttered: birdsong, wind gusts, rustling vegetation, the occasional aircraft overhead. Sorting turbine noise from everything else, across tens of thousands of minutes of audio sampled at tens of kilohertz, has remained an unsolved bottleneck — and without a reliable way to do it, studies of how turbine sound travels and affects nearby residents stay incomplete.
The noise identification problem wind energy researchers face
Wind turbines are moving closer to residential areas as renewable energy capacity expands worldwide. That proximity makes understanding turbine noise behavior — not just its volume, but its specific character — a pressing scientific and regulatory concern. Long-term field measurements capture how turbine noise behaves across seasons, weather patterns, and operating conditions. But those recordings don’t arrive pre-sorted. A microphone positioned near a turbine picks up everything: birdsong, wind gusting through vegetation, the occasional aircraft, and the turbine itself, all layered together across months of continuous audio.
Existing approaches struggle with this complexity. Statistical methods based on sound level percentiles can misclassify other sources that happen to behave similarly in the frequency domain — aircraft overflights and wind-induced vegetation noise are especially prone to causing false positives. Until now, no validated, computationally practical framework existed to systematically identify turbine noise and its specific acoustic components within large, messy datasets.
Tonal events clustered around approximately 60, 240, and 520 Hz — proportions consistent with gearbox rotation speed, pointing toward a mechanical drivetrain origin.
How the two-stage framework works
The new framework tackles the problem in two sequential passes. Stage 1 combines statistical preselection with a physics-based signal analysis. The preselection filters out minutes with incomplete data, periods when neighboring turbines are running, and conditions where wind speed exceeds 15 m/s — a threshold above which wind-induced microphone and vegetation noise becomes a serious contaminant. What survives then undergoes a more demanding test: the algorithm analyzes amplitude modulation in third-octave frequency bands between 63 and 2,000 Hz, sampled at 20 Hz.
The algorithm looks for periodic fluctuations in sound level, then checks whether their frequencies match the blade-passing harmonics calculated from SCADA rotor-speed data. A period gets flagged as turbine-dominated only when a coherent harmonic pattern appears across at least five of seven mid-frequency bands and persists across multiple consecutive minutes. Stage 2 identifies three specific noise components: rotor-induced amplitude modulation (AM), high-frequency whistling noise linked to a single blade, and tonal mechanical components assessed following IEC 61400-11. Processing the full one-month dataset of 43,200 minutes took roughly 12 hours.
Validating the algorithm against human listeners
To check whether the algorithm’s classifications matched human perception, the researchers ran a structured listening test. Fifteen participants — all employees of the Institute of Structural Analysis, three of them experienced in wind turbine acoustics — evaluated approximately 1,200 ten-second audio segments. Their responses were aggregated into 166 classified minutes that served as a perceptual reference. Participants sorted sounds into six categories: rotor noise, whistling, tonality, wind-induced noise, bird calls, and other sources.
Intrarater reliability was good to excellent, with a mean Jaccard index of 0.87. Agreement between different listeners averaged 0.56 and varied considerably by category. Tonality proved the most subjective, with notably low interrater consistency. The researchers are careful to describe this reference as consensus-based rather than ground truth.
What the validation results revealed
Stage 1 performed with near-perfect precision. Of all the minutes the algorithm flagged as turbine-dominated, 99% matched what listeners identified as turbine noise. Recall reached 0.96, meaning the algorithm caught the vast majority of genuinely turbine-dominated periods. The handful of false negatives occurred mainly when turbine sound competed with loud bird calls or wind-induced noise — by design, since the framework aims to identify acoustically dominant turbine noise, not every moment when turbine sound is merely audible.
Whistling noise detection achieved high precision (0.78) but lower recall (0.42), with most false positives involving high-frequency bird vocalizations. Tonal component detection showed the inverse pattern: high recall (0.94) but moderate precision (0.46), with false positives clustering near the audibility threshold in low frequencies. The researchers interpret these discrepancies as reflecting the limits of subjective perception rather than flaws in the algorithm itself.
Patterns uncovered in a full month of turbine noise data
Applied to the complete one-month dataset, the framework identified dominant turbine noise in 10,767 of 43,200 minutes — roughly 24.9% of total measurement time. AM detection rates peaked in the 160–800 Hz mid-frequency range, consistent with near-field aerodynamic trailing-edge noise and aligned with the frequency range the Institute of Acoustics recommends for near-field AM evaluation. Detection rates rose with rotor speed and atmospheric stability, and fell at high wind speeds as wind-induced masking increased.
Tonal events clustered around approximately 60, 240, and 520 Hz — proportions consistent with gearbox rotation speed, pointing toward a mechanical drivetrain origin. The framework, listening-test platform, and anonymized audio signals are all publicly available under FAIR data principles.
What comes next for wind turbine noise research
This framework is already slated for deployment across four long-term measurement campaigns at two sites, with microphone distances extending to roughly 1,500 meters. At those distances, atmospheric and ground effects on sound propagation become significant. The current validation comes from flat terrain in northern Germany. Forested areas, complex topography, and different climate regions will introduce scattering, shielding, and altered emission characteristics — all of which may require parameter adjustments. The modular design is intended to accommodate exactly those kinds of extensions.
There’s also room to expand what the framework detects. Tonal amplitude modulation — a distinct and often more annoying form of wind turbine noise — is one candidate for a future module. As turbine designs evolve and installation sites diversify, having a validated, adaptable tool for parsing long-term acoustic records will matter more, not less.
The full study is available here: Könecke, S., Jonscher, C., Bohne, T., and Rolfes, R.: A two-stage framework for identifying and characterizing wind turbine noise data and its validation by listening tests, Wind Energ. Sci., 11, 1771–1789, https://doi.org/10.5194/wes-11-1771-2026, 2026.
Carlos is an engineer with strong expertise in technical and industrial topics. He previously worked at international companies such as Siemens and is multilingual.
