A Design of Seismometer Anomaly Detection System Based on Frequency-domain Features

Although having an important role in seismic signal monitoring, seismometers are prone to several fault scenarios caused by aged electrical components or faulty pendulum mechanisms, which may affect the measurement quality. Some of these faults generate non-seismic noises which result in anomalous power spectra in the frequency-domain. To ensure accurate measurements, seismometers' health condition must be monitored based on their measurement data. This encouraged us to build a design of anomaly detection system based on the frequency-domain features. These features were extracted by calculating the deviation between seismic power spectra and Peterson's ambient noise models. The performances of one-class SVM, local outlier factor, one-class autoencoder, and isolation forest models were compared in this study. The one-class SVM came out with an outstanding performance, proved by an accuracy of 0.994 when detecting anomalies in the test data. Lastly, the proposed system may become an effective option to be implemented as a part of regular maintenance procedure in seismic observation networks.