機器學習和低價位地震儀於現地型地震預警之應用

Abstract

地震預警系統為地震減災的方法之一。地震預警系統，分為區域型及現地型地震預警系統。區域型地震預警系統，收集並分析近震央測站的資料，提供詳細的地震資訊。而現地型地震預警，藉由分析P波後到達幾秒內波形，對近震央區域提供快速的地震警報。本研究主要探討現地型地震預警的功效、原理，並嘗試用機器學習提升系統效能。臺灣近年所使用的現地型地震預警系統，利用P波到達前幾秒的最大位移量(Peak displacement, Pd)是否超過0.35cm做為觸發門檻，有著相當好的成效，對臺灣地震減災提供了巨大的貢獻。但是，在資料處理的過程中，不同濾波器的參數使用會影響資料，造成Pd取值上的差異，進而影響系統效能及判識。此外，現地型地震預警系統雖然有出色的低誤報率，但是由於積分過程會壓低高頻訊號，可能會低估原加速度值達災害事件門檻的近震事件，導致漏報發生。而機器學習能分析並學習資料的特徵，不需要進行任何資料預處理，就可以正確判識資料。基於機器學習的優勢，本研究選取42個島內近震事件之原始加速度資料訓練機器學習的模型後，針對2018年芮氏規模6.26花蓮地震及2019年芮氏規模6.3秀林地震進行測試，並比較本研究與傳統Pd方法之效能。結果顯示，只利用原始的加速度訊號，也可以得到傳統Pd方法的低誤報率，並且大幅降低漏報率。運用機器學習的技術，得以分析原始訊號中的特徵，並提供了一種更穩定且可信的地震預警模型。

Earthquake early warning (EEW) plays an important role in earthquake hazard mitigation. There are two types of EEW system, regional and onsite. Onsite EEW systems analyze initial part of the seismic waves from the P-waves to predict later ground motion from the S-waves and surface waves. In recent years, an onsite EEW method based on fixed peak displacement (Pd) threshold is developed. Although the method consistently provides effective warnings in Taiwan, several studies suggest that its strong filter dependence might introduce extra biases to the system. Also, the fixed Pd threshold suffers from the inevitable trade-off between a false alarm and a missed alarm. In order to overcome the abovementioned problems of fixed Pd threshold method, we utilize techniques in machine learning and develop a new method of onsite early warning. Owing to the property that convoluted-neural-network (CNN) will automatic sampling on different frequencies, the unfiltered seismic signal itself is sufficient to derive a warning threshold. As an example, we collect 42 medium to large inland earthquakes in Taiwan. We compare the performance between the proposed method and the fixed Pd threshold method on 2018 ML 6.26 Hualien and 2019 ML 6.3 Xiulin earthquakes. The result shows that the proposed method not only outperforms in stability, also in missed alarm rate and false alarm rate. The proposed method can provide not only significant improvements for onsite EEW but also a window into the initial P-waves. Studying the trained models might even reveal the hidden indicators inside the initial P-waves.