基於單類別分類之構造長微震偵測架構設計 Tectonic tremor detection framework designbased on one-class classification

Date
2021
Authors
吳宇翔
Wu, Yu-Siang
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
在臺灣,自發型構造長微震(以下簡稱「長微震」)之好發區域為中央山脈南段,具有(1)持續時間長,可達數分鐘至數小時、(2)不具明顯可見之體波、(3)能量富集於2至8 Hz間,並可在數十公里遠的測站有幾乎一致的到時特性,而偵測手段仰賴多測站的包絡化波形進行互相關係數與測站間到時差。前人研究也發現,在臺灣進行長微震偵測時,較吵雜的背景噪訊與短時間密集發生的區域地震(震央距50-200公里)容易與長微震波形混淆,使最終的長微震目錄底定必須經過人工目視,較為耗時且涉及主觀成分。為探索以機器學習進行地震與長微震自動分類的可能,本研究以k-最近鄰居法搭配29項特徵對2016年間5,796筆區域地震與6,746筆長微震事件進行分類,搭配循序向前特徵選取法(Sequential Forward Feature Selection)達到96.4至99.1 %分類率,初步證明運用機器學習於長微震分類上之可行性。然而訓練多類別分類器必須針對所有類別進行定義、抽樣與標籤化,難以實現於連續偵測。本研究進一步以單類別分類器支援向量資料描述(Support Vector Data Description),設計長微震連續偵測架構,其優勢在於只需要長微震資料進行訓練,而不需針對大量類別進行處理。藉由設立多測站投票制度與持續時間門檻以及使用2016年1月1日至7月18日長微震事件進行訓練,本研究成功於2016年7月19日至9月10日,使用三個測站偵測出共132,240秒長微震。當提升測站數至九站,只使用水平分量於單站決策並在多站投票時以各站訊噪比為權重,偵測出總計10,620秒的長微震事件,但經目視後保留之事件比例,從使用三站的5.8 %提升至九站的31.6 %,證實了應用單類別分類於多站長微震偵測的可行性。
In Taiwan, ambient tremors are found to locate underneath a mountain in Southern Central Range, a core of collisional mountain belt. Given that the ambient tremors represent the aseismic slip process at greater depth where no seismicity is present, it is crucial to monitor their activity. To identify tremors, the similarity and time lapse of the arrival tremor bursts from multiple stations are oftentimes demanded, while manual checks of multi-station waveforms are practiced, to exclude loud noise and swarms of regional earthquakes (epicentral distance of 50-200 km). Here we attempt to exploring if the advances of machine learning (ML) techniques enable an automatic search for patterns to discriminate tremor from regional earthquakes. Using k-Nearest Neighbor (k-NN) classifier with 29 features including band-passed filtered energy, frequency spectrum and spectrogram, Liu et al. (2019) successfully separated tremor from local earthquakes and noise at high accuracy of 86.6-98.8 %, showing that the possibility of applying ML technique to separate tremors from other types of seismic signals is feasible. However, they also found that the multi-class classification approach is not robust in practical, especially in the real-time monitoring using continuous data. This is mainly due to the fact that in the continuous data, many other types of signals exist without being labeled. In this study, we used two-class and one-class classifier to demonstrate if tremor can be separated from regional earthquakes and the possibility of continuous tremor detection.Using 5,796 tremor, 6,746 regional earthquake, and 441,887 noise data collected in 2016, we found the two-class k-NN classifier allows the high classification rate (CR) of 91.8-95.8 %. However, in real-time tremor monitoring using continuous data, the multi-class classification may not be practical due to the under-sampled data. We further demonstrate the performance of one-class Support Vector Data Description classifier that does not need a collection of other classes’samples. The resulting CR of 80.8 % indicates the capability of one-class classifier on real-time detection of tremor in Taiwan.
Description
Keywords
臺灣, 構造長微震, 機器學習, 自動偵測, Taiwan, tectonic tremor, machine learning, real-time monitoring, classifier
Citation
Collections