Stage 2 examines the 1 min periods identified in Stage 1 in order to determine specific WTN components. These include AM, whistling noise, and tonal components.

For validation of the proposed framework, an assessment based on 1 min sound recordings was required. The recordings were selected under reference conditions expected to yield representative WTN characteristics. These conditions were derived by systematically categorizing the measurement data set and are mostly aligned with the preselection criteria in Sect. . They comprise rotor speeds above 11 rpm, wind speeds below 11 m s⁻¹, and downwind or downwind–crosswind wind directions. The A-weighted equivalent sound pressure level was restricted to 25≤LA,eq,1min⁡≤60 dB(A), low temporal variability of percentile sound levels was required, and stable stratification was assumed (α>0.2, where α describes the vertical wind shear according to the power-law wind profile). These reference conditions formed the baseline scenario for all subsequent systematic parameter variations, which are summarized in Table . For each parameter combination, 10 data sets of 1 min duration were randomly selected, if available, with each data set being used only once. In total, the listening-test data set comprised 194 individual 1 min recordings, including approximately 10 % blinded repetitions to assess intrasubject consistency, i.e. the extent to which the same participant provided consistent responses when evaluating two identical recordings. Since the perceptual evaluation of continuous 1 min recordings is cognitively demanding, each recording was subdivided into six consecutive 10 s segments, which were evaluated individually. For each 10 s audio segment, participants could select nine response categories: uncertain, no WTN, WTN in combination with disturbance noise (wind-induced noise, birds, or other sources), and specific WTN components (rotor noise, tonality, machine noise, or whistling noise). The playback and evaluation were performed using a graphical user interface (GUI) implemented in MATLAB (Fig.  in the Appendix).

Prior to the test, WTN was defined explicitly as periodic aerodynamic noise generated by the rotating rotor blades. Machine noise, such as sounds caused by nacelle yaw movements, was explicitly distinguished from tonal components. Participants were also able to provide free-text comments, for example to indicate partial masking of WTN by wind noise.

The listening test comprised three standardized phases. In the introduction phase, participants received a guided presentation of the test environment and listened to 16 representative audio examples illustrating typical WTN characteristics and common extraneous noise sources. The training phase included a test run with comparison to a reference solution; in the case of deviations, audio segments were replayed, and open questions were discussed. In the main test phase, participants evaluated randomly ordered audio segments. After every 3 min of playback, a neutral audio signal (elevator music) was presented to neutralize the hearing. Breaks were optional after 15 min and mandatory after 60 min.

The entire listening test comprised 3 h and 14 min of pure audio material and was conducted by 15 participants, all employees of the Institute of Structural Analysis. Three of them work directly in the field of wind turbine acoustics and are experienced in WTN. The remaining test participants work in other wind-energy-related fields. All participants used identical headphones, namely Real Blue NC from Teufel.

To validate the developed framework, a reference data set is required that provides a reliable classification of each 1 min segment. Since the listening test was based on 10 s audio segments, the listener responses are aggregated to a minute-level consensus.

For each participant and each category, six binary evaluations corresponding to the consecutive 10 s segments of a 1 min period are available. To derive a clear minute-level classification for each category and participant, these six judgements are aggregated using a strict majority rule. Based on the number of positively rated 10 s segments s∈0,1,2,3,4,5,6, the minute-level classification is defined as follows:

category present, if s≥4;

If none of the evaluated categories reaches the threshold for being classified as present, a minute is defined as unclear at the consensus level.

To validate the results from Stage 1, i.e. the identification of dominant WTN, the minute-wise consensus label from the listening test for the category “WTN, rotor” is used as a reference classification.

A period of 10 min was classified as “uncertain” in the listening test and is excluded from validation, as there is no clear reference for these periods. Minutes labelled “no WTN” (label 2) are considered valid negative examples. Validation is performed on a minute-by-minute basis by comparing the algorithmic classification with the reference labels using standard classification performance metrics. The results are summarized in Table .

For the 166 valid minutes, the algorithmic classification and the listening test agree very well (TP: 140, TN: 19). The number of false positives is extremely low (FP: 1). This misclassification occurs at a low rotor speed of 7.99 rpm and was discussed as a special case in the plausibility check (Sect. , Fig. ). A possible explanation is the limited temporal resolution of the available SCADA data, as short-term operating conditions cannot be captured by 10 min averages. Most misclassifications correspond to false negatives (FNs: 6). In three out of these six cases, the WTN occurs together with other noise categories, such as wind-induced noise and bird calls. From a methodological point of view, this behaviour is intended. The aim of Stage 1 is not to detect all perceptible WTN, but to identify acoustically dominant WTN. In situations with a high level of background noise, WTN is therefore intentionally not classified as dominant. The high precision (0.99) shows that a positive methodological classification nearly consistently corresponds to the perception of a WTN in the listening test. False positive decisions therefore only occur in exceptional cases. The recall (0.96) illustrates that the majority of actually perceived wind turbine noises are correctly identified, while the remaining false negatives occur primarily in acoustically complex situations. The high specificity (0.95) also confirms that time periods without WTN are reliably recognized as such.

In addition to a global decision on the dominance of WTN, Stage 1 provides a band-resolved assessment of whether WTN is identified as dominant in individual third-octave bands. As part of the validation, the following analysis examines whether this band-specific methodological assessment is consistent with the disturbing noises identified in the listening test. Since the listening test is not frequency resolved and does not provide an explicit assessment of individual third-octave bands, it cannot serve as direct reference for validating the band-specific results. Instead, it is used as an independent reference to contextually verify the Stage 1 results and assess their physical plausibility. Minutes are grouped according to the presence of specific interfering noise sources identified in the listening test, and the band-wise Stage 1 results are compared across these groups.

Wind-induced noise (category 3) and birds (category 4) are grouped exclusively in order to isolate the influence of a single source of interference. This means that only minutes in which both WTN and the respective interfering source were identified are included in the respective group. A strictly defined reference group (“pure WTN”) is constructed from minutes in which only WTN was identified and no additional disturbing noise was reported. This reference represents WTN under minimally disturbed acoustic conditions and enables a comparison with specific interference scenarios.

Figure  illustrates the results in terms of the proportion of non-detected third-octave bands, i.e. the fraction of bands in which Stage 1 does not classify WTN as dominant. The following results should not be understood as a major validation of the band-wise assessment, but rather as a supplementary plausibility and consistency check. For minutes with wind-induced noise (category 3, n = 7), an increase in band-wise non-validity is observed compared to the pure WTN reference (n=63). In the third-octave bands between 63 and 125 Hz, the proportion of non-detected minutes increases strongly, reaching differences of up to 0.85 in the 63 Hz band. The effect is also noticeable in higher third-octave bands (1000–2000 Hz). The resulting pattern indicates a systematic influence of wind-induced noise on the band-wise dominance assessment. Wind-induced noise can occur in both the low- and the high-frequency ranges, depending on its source. Wind-induced noise at the microphone primarily affects low frequencies, while wind-induced noise generated by vegetation can contribute to higher-frequency components.

For bird sounds (category 4, n = 64), the effects are considerably weaker than for wind-induced noise. Although the band-wise non-detection is generally higher than in the pure WTN reference, no pronounced frequency-specific pattern is observed. Deviations occur in both low and high third-octave bands, but their magnitude remains well below that observed for wind-induced noise. Since bird calls predominantly occur at frequencies above approximately 2 kHz, no exclusively high-frequency pattern is discernible in the present analysis. However, the presence of deviations in both low- and high-frequency bands during bird-dominated minutes indicates that additional low-frequency components contribute to the observed non-detection. These findings suggest that weak low-frequency wind-induced noise components are not always classified in the listening test. Particularly when they are weak, such components can apparently be perceived as part of the WTN or overheard. In this context, it can also be assumed that even in the reference minutes (pure WTN), residual low-frequency noise components may be present but are not perceived as independent noise.

As in the previous validation steps, the minute-wise consensus classification from the listening test is used as a reference classification for validating the detection of whistling noise. Validation is performed on a minute-by-minute basis. With respect to the methodological results, only the 2000 Hz third-octave band is considered for validation.

Below 1600 Hz, no whistling noise was detected by the framework during the listening-test periods. Although individual detections occur in the 1600 Hz band, these are predominantly observed concurrently with detections at 2000 Hz. In addition, the corresponding C values at 1600 Hz are usually only slightly above the threshold value (C=4), while the 2000 Hz band shows higher C values. Consequently, whistling noise detection is validated exclusively for the 2000 Hz band.

The validation results, expressed in terms of derived classification metrics, are summarized in Table . In the listening test, 59 out of 166 min was rated as whistling noise by consensus. Using the proposed method, whistling noise was detected in 31 out of 166 min. The validation results show a high precision of 0.78, indicating that the detections are generally correct. The recall is in a lower range of 0.42.

Of the 59 min classified as whistling noise in the listening test, 35 min is not detected by the method (FN: 35). A detailed analysis of these false negatives shows that 4 min is already excluded in Stage 1 because no dominant WTN is identified. In 2 additional minutes, Stage 1 is fulfilled, but no dominant WTN is detected in the 2000 Hz band. In 10 min, the single-blade frequency p1 is not among the four strongest modulation frequencies determined in Stage 2. In the remaining 18 min, both Stage 1 and the p1 condition are satisfied, but the prominence ratio remains below the defined threshold (C<4).

A total of 7 min is incorrectly classified as positive (FP: 7), which corresponds to 4.0 % of all listening-test minutes or 21.9 % of the method-detected whistling noise events. Further analysis of the false positive data shows a clear correlation with the listening category “birds”. While this category accounts for 45 % of the total data set, it is likely overrepresented in the false positives at 71 %. In 5 of the 7 false positive minutes, bird song was clearly annotated. Due to the high-frequency spectral content of bird vocalizations, perceptual discrimination between bird sounds and whistling noise can be challenging. Re-listening to these minutes confirms this statement. In the remaining 2 false positive minutes, whistling noise is only weakly perceptible. Further investigations showed no consistent correlation with the prominence ratio C or other hearing categories.

The detection of tonal components based on IEC 61400-11 can be validated directly using the minute-wise consensus classification of listening-test category 7 (“WTN – tonality”), which serves as the reference classification in the following analysis. It should be noted that the listening test did not include an explicit assessment of tonal frequencies or audibility levels. Instead, tonality was evaluated as binary, i.e. with respect to its presence or absence. The validation results are summarized in Table . A total of 17 out of 166 min (i.e. 10.2 %) was rated as tonal in the listening test. Thus, the reference data set is comparatively low.

Out of the 17 min classified as tonal in the listening test, 16 min is correctly identified by the method, resulting in a high recall of 0.94. This indicates that almost all perceptually identified tonal events are detected by the algorithm. At the same time, the majority of minutes not classified as tonal in the listening test are correctly recognized as non-tonal (TN: 130), corresponding to a specificity of 0.87. However, tonal components are detected methodologically in 19 min that was not classified as tonal in the listening test, resulting in a moderate precision of 0.46.

In the only false negative minute, tonal components are detected in the third-octave bands with centre frequencies of 250 and 500 Hz. The corresponding frequency-corrected audibility values are dLA=-9.8 dB (250 Hz) and dLA=-4.2 dB (500 Hz). While the value at 500 Hz lies close to the audibility threshold of −3 dB, the value at 250 Hz is clearly below this threshold. Re-listening to this minute showed that the tonal components are only marginally perceptible, indicating a borderline perceptual case. With 19 min, the number of false positives is significantly higher so that correlations between tonality detection, detected sound frequencies, and environmental and perception conditions are considered. In 12 of these minutes, tonal components are detected in the low-frequency range (3 min at fc=63 Hz, 6 min at fc=80 Hz, and 1 min at fc=100 Hz). The majority of these detected tones have frequency-corrected audibility values close to the audibility threshold (dLA between −3 and 3 dB) and are therefore only weakly audible. One exception is a single event in the 80 Hz third-octave band with a high audibility value of dLA=12.66 dB. It is therefore plausible that low-frequency tonal components with low to moderate audibility were not consistently recognized by the participants in the listening test. In this frequency range, wind-induced noise cannot be excluded as a confounding factor. Furthermore, bird sounds occur frequently in the false positive minutes, indicating context-dependent perception by the test persons.

In the mid- and high-frequency third-octave bands (fc=250 Hz: 3 FP, fc=400 Hz: 1 FP, fc=500 Hz: 5 FP), false positives occur almost exclusively near the audibility threshold. The detected tonal components are therefore not sufficiently audible to be classified as dominant tonality by the majority of test participants in the listening test. In these cases, the low audibility provides a plausible explanation for the discrepancy between algorithmic detection and listening-test classification.

The false positives observed can be interpreted less as a shortcoming of the method and more as a consequence of the subjective and context-dependent evaluation in the listening test. The method evaluates tonality strictly according to physical-psychoacoustic criteria, whereas the listening test reflects a dominance-based perceptual decision in the presence of competing sound sources. The reliability analyses in Sect.  already showed that tonality exhibits comparatively low interrater consistency. To conclude, the deviations can primarily be explained by perception effects.

The amount of available data and the percentage of minutes excluded in each step of Stage 1 are summarized in Table . Initially, 17.29 % of the measurement data exhibit incomplete recordings, with parts of the acoustic time series missing. These minutes are excluded from further analysis.

Preselection has by far the largest impact on data reduction. After applying this step, 11 407 min remains, corresponding to 26.4 % of the initially available 43 200 min. A major reason for this high exclusion rate is the requirement that neighbouring wind turbines must be inactive to enable the evaluation of an individual turbine. These results demonstrate that preselection is an effective filtering step, removing a large fraction of data that is unlikely to contain dominant noise from one WT prior to the computationally intensive signal analysis. As a result, only time periods with potentially dominant WTN are considered in the subsequent, significantly more complex analysis steps. Compared to preselection, the signal pattern conditions applied in Stage 1 have a smaller impact on further data reduction. The combined application of the harmonic pattern and spectral occurrence conditions results in the exclusion of only 634 additional minutes, corresponding to 5.56 % of the initially data. With only 6 min excluded, the temporal occurrence condition has a negligible impact on the data set considered here. In total, dominant noise from the wind turbine under investigation is detected in 10 767 of 43 200 min, corresponding to 24.9 % of the total measurement time.

The detection rates for the individual third-octave bands, based on the 10 767 WTN-dominated minutes, are shown in Fig. . The detection rate exhibits a frequency dependence, with a pronounced maximum in the mid-frequency range and a decrease towards both lower and higher third-octave bands. The reduced detection rate in the 63 Hz third-octave band (detection rate of 0.75) is due to wind-induced noise at the microphone. A decreasing detection rate is also observed at frequencies above 1 kHz. This is primarily attributed to wind-induced vegetation noise in the vicinity of the measurement position. Due to the relatively short distance of 178 m between the microphone and the wind turbine, atmospheric absorption is expected to have only a minor influence overall. Under standard conditions (20 °C, 50 % relative humidity), attenuation at 2 kHz is about 1.4 dB per 100 m. Atmospheric absorption may reduce the detection at 2 kHz, while its effect is not pronounced at lower frequencies. However, at greater propagation distances and across a wider range of atmospheric conditions, particularly under low humidity and higher temperatures, atmospheric absorption becomes more significant.

The highest detection rates, ranging from 0.96 to 0.995, occur in the mid-frequency third-octave bands between 200 and 800 Hz. This frequency range corresponds to the A-weighted dominant range of wind turbine emissions. It should also be noted that, due to the spectral occurrence condition applied, only minutes in which at least five of seven third-octave bands between 200 and 800 Hz show a simultaneous detection are classified as WTN.

The Stage 1 detection results are subsequently analysed with respect to environmental and turbine operating conditions, namely rotor speed, wind speed, propagation direction, and atmospheric stability. The corresponding detection rates are shown in Fig.  and are defined as the fraction of minutes classified as wind turbine noise dominant relative to the total number of available minutes in the respective class, considering only periods in which the turbine under investigation is operating while all other turbines are inactive. The detection rate increases as the rotor rotational speed increases. This behaviour is physically plausible due to increasing emission levels with rotor speed and hence higher signal-to-noise ratio. Regarding wind speed, the detection rate exhibits a maximum in the range of approximately 7–11 m s⁻¹. At lower wind speeds, rotor speeds are reduced, leading to lower detection rates, consistent with the correlation described above. At higher wind speeds, however, the detection rate decreases significantly, which is consistent with increasing wind-induced noise and enhanced masking of wind turbine sound. Furthermore, the detection rate increases with greater atmospheric stability. This is also physically plausible, as stable stratification is associated with less atmospheric turbulence and usually occurs at night. As a result, background noise is lower. With respect to the propagation direction, lower detection rates in downwind conditions are consistent with reduced AM characteristics reported in . Since the method used here is based on the correlation between the BPF and detected harmonics, and thus evaluates an AM-consistent feature, a possible explanation for the reduced detection rate is a lower incidence of amplitude-modulated characteristics in the downwind direction.