Machine learning for data-driven pattern recognition of seismic wind turbine emissions

SUMMARY Seismic emissions from wind turbines (WTs) depend on the rotation of the WT blades and the wind direction-dependent movement of the WT. Mechanical coupling between the WT foundation and the subsurface generates complex seismic wavefields, making it challenging to manually separate the contributions of different signal sources, thus complicating data labelling. We address this challenge by applying unsupervised machine learning techniques that do not require labelled data. Our analysis focuses on seismic WT emissions recorded near Wind Farm Tegelberg in the eastern Swabian Alb, Southwest Germany. Specifically, we extract time-averaged wavelet features by temporal averaging the wavelet transformation of the continuous three-component seismic data and subsequently apply the clustering algorithm Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN). The resulting clusters not only capture the variations in the WT rotation rate but also reveal a clear dependence on wind direction, associated with the radiation pattern of different surface waves. Our results demonstrate the potential of HDBSCAN to uncover meaningful, source-related patterns in continuous seismic records.