Deep learning model for co-seismic response detection of fixed-point deformation data

Co-seismic response identification of fixed-point deformation data currently relies on manual selection, and no automatic co-seismic response detection method has yet been applied. This study proposes the first deep learning model for the co-seismic response detection of fixed-point deformation data in China, which is used to detect co-seismic response signals quickly and accurately on the second dataset of a single vertical pendulum broadband tiltmeter. The model was constructed using the transfer learning technique; it introduces three representative pre-trained models for earthquake detection in seismic data as feature extractors, migrates their knowledge and capabilities of earthquake detection in seismic data to fixed-point deformation data, and then designs and adapts supporting data converters and classifiers. Tests on real observational data showed that the model provided a good detection performance. The application of continuous data from Jixian station proved that the model was not only capable of detecting all the co-seismic response events recorded manually, but it also found events that were not recognized manually, with an accuracy rate of no less than 75%. Compared with traditional manual processing, the detection efficiency, detection capability, and consistency are greatly improved using this model.