Seismologist noticed are worsening of station quality after the installation of wind turbines (WTs) close to seismological stations. Since WTs and seismological stations are installed mostly in areas with low population density, both are looking for solutions to diminish this conflict.
For this, we tested different denoising techniques at affected seismological stations to reduce or to eliminate the disturbing WT signal from the seismological data. Usually, spectral filtering is used to suppress noise in seismic data processing. However, this approach is not effective when noise and signal have overlapping frequency bands which is the case for WT noise. First, we applied a nonlinear thresholding function on our data. This method leads to good results when the event can be already seen in the raw data but it fails when the event is fully covered by noise. As a second method, we used a denoising autoencoder (DAE), which learns a sparse representation from time-frequency coefficients, and maps from there to output masking functions for signal and noise. The DAE is more time consuming in comparison to the nonlinear thresholding function but when the convolutional neural network is trained, using an adequate training dataset, it can also be applied instantly on the raw data. The DAE distinguishes between signal and noise and we are able to correct the seismograms from most of the disturbing noise signals.