Earthquake signal dynamic threshold detection by fusing multiple waveform features in sparse variational Gaussian process regression

Aiming at the problems of the low detection rate of earthquake signals and sensitivity to the detection threshold, the dynamic threshold detection algorithm of sparse variational Gaussian process regression (SV-GPR) fusing multiple waveform features is constructed. (i) Four physically interpretable features—the short-term average over the long-term average (STA/LTA) waveform amplitude feature, the STA/LTA waveform ratio feature, the spectral feature, and the intrinsic mode function energy feature—are selected to analyze the historical seismic data in the specific region. (ii) The earthquake signal detection prediction model is constructed related to distance by Gaussian process regression (GPR), and the form of its sparse variational approximation is given. (iii) The Bayesian joint distribution algorithm is designed to fuse the detection results of the four features to obtain the prediction model with dynamic thresholds, and the standard interval confidence method is proposed to evaluate the prediction model performance of the proposed SV-GPR algorithm. The detection accuracy and computational efficiency of the proposed SV-GPR algorithm are validated with the seismic dataset from the Incorporated Research Institutions for Seismology. In terms of detection performance, the SV-GPR algorithm improves the precision by 13.99% and the recall by 26.62% compared with the STA/LTA algorithm. In terms of computational efficiency, the proposed SV-GPR algorithm improves the computational efficiency by 80% compared with the GPR algorithm with a loss of 2% detection accuracy.