Quality control of seismic data based on convolutional neural network

Installing more seismic stations may result in improving the capability of earthquake monitoring and shortening the time to report the occurrence of earthquakes or deliver public earthquake warnings. However, accordingly, it becomes difficult to assess the condition of seismic instrument. The goal of this study is to develop an automated method for assessing the quality of seismic data, which is based on power spectral densities (PSD) of one-hour waveform data. We collected 10,309 PSDs of broadband seismometers and 4,452 PSDs of accelerometers recorded from 2016, 2017 and 2019, and used them as the input of the convolutional neural network (CNN), a class of deep learning. Two deep CNNs for broadband seismometer and accelerometer were trained to automatically determine the condition of seismic data: normal and abnormal conditions. We find that the condition of seismic data determined by the CNNs have an accuracy of 99.9% and they can successfully determine the condition from PSDs of 15-minute waveforms. The outstanding performance of the trained models indicates that this can be a very effective tool for assessing the condition of seismic instrument.