應用機器學習方法於日本四國長微震訊號之分段特徵

Abstract

構造長微震為慢地震族群中的一個類別，此特殊震動訊號貌似雜訊，其特性為（1）振幅微弱，無明顯體波波相；（2）主頻落在2至8 Hz之間；（3）能量持續時間長，從數分鐘到數天不等；並且（4）在相距數十公里的測站具有相近的到時，若不以多個測站之波形判斷，則很難將長微震與雜訊區分。本研究將探討應用機器學習方法於自動化分類，對於跨區段的長微震訊號是否可行？而分類後對於訊號的本質，是否能夠進一步得到能反映這些訊號的物理背景？著重對於訊號本質並且配合特徵及物理意義討論，相較一般觀測數值之計算，預期得知蘊藏在訊號中的細節。使用監督式學習（supervised learning），應用k­最近鄰居法（k-NN）、向前特徵選取法（sequential forward feature selection），並且配合費雪分數（Fisher score）的計算以整合有效特徵。由於日本四國長微震之時空分佈特徵清晰、分段明顯，本研究訓練分類器以區分四國四個不同子區域（A-D）的長微震訊號。選用Annoura et al. (2016) 目錄從 2014 年 6 月 1 日到 2015 年 3 月 31 日的事件，將目錄中不同子區域的長微震進行波形切段，時間長度為 60 秒，得到10000至31000 筆多筆資料，以形成數據集合。使用每筆標籤化之60 秒數據，針對 29 個地震特徵應用 k-NN 分類器來研究其分類性能。除了針對訊號於不同震源區傳遞至上方測站接收，也固定了單一震源區探討傳遞至上方測站以及其他區段測站之訊號，將不同區段的訊號特性分類比較，進行了5個試驗，共18個子測試，高分類率（90 % 以上）的結果表明應用k-NN 來區分不同區段長微震的潛力。利用費雪判別準則綜合P值（(p-value）檢定，發現波形資料在自相關函數以及離散傅立葉轉換計算後之特徵具有較高的費雪分數，在分離不同區段長微震中擔任著重要的角色。此外，由於長微震發生的物理機制與構造環境不同，而高分離性的特徵也能進一步加入解釋及探討。本研究最終成功地分離了四國不同區段的長微震訊號，並篩選出於不同子區域有效的震源差異分類特徵，表示了應用機器學習區分震源特性的可能。

Tectonic tremor is long-lasting (several minutes to days) and noise-like seismic signal with the dominant frequency of 2 to 8 Hz. It is generally identified by waveform similarity in the envelope with nearly the same arrival at different stations. Without a display of waveforms from multiple stations, discriminating tremors from noises is difficult. How much the tremor signals can tell us about the source properties and regional differences in physical environments that host tremors? Applying machine learning, how can we demonstrate the segmentation based on signal characteristics? To answer these questions, this study attempts to explore the signal characteristics of tremors in areas that are naturally separable and different from each other. We used the trained classifier for discriminating tremors in four different sub-areas of Shikoku (A-D) due to the clear segmentation of tremor distribution and characteristics in space and time (Kano et al., 2018). The tremor events in the Shikoku region of southwest Japan were from the catalog by Annoura et al. (2016) from 1 June in 2014 to 31 March in 2015. To prepare the labeled events composed of the largest number of data, tremors in the catalogs were segmented into 10000-31000 signals with a length of 60 seconds. Using the 60-s-labeled data, we applied k-NN classifier on 29 seismic features to investigate the classification performance. Based on pairwise comparisons among a set of tremors and stations at these sub-areas, the high classification rate (higher than 90 %) suggests the potential of applying k-NN on tremor discrimination in different areas. Under Fisher's class separability criterion, we found that the optimal features with high Fisher scores on the spectral contents and energy concentration play important roles in separating the tremor signals. The high separability is somewhat expected due to the different physical and tectonic environments for tremor generation. In this study, we successfully separated the tremors in the distinct segments of Shikoku. Investigating the variations in signal characteristics reveals the possibility of differentiating the source properties by using machine learning.