利用深度學習解析震源機制解以及完善 AI 地震觀測平台 SeisBlue

Abstract

應對密集地震網和地震資料量劇增的重大挑戰，本研究注重在應用深度學習技術來提高地震資料處理的自動化程度和效率，尤其是在規模大於 0.8 以上的微震觀測領域的發展上。自 2018 年以來，台灣大學構造地震研究室開發 SeisBlue 系統並在 2020 年開始成功應用在多個案例，達到區域性的近即時地震定位。 本研究致力於開發更完善的第三代系統，目標著重在系統擴展、強化和整合。新開發一個以卷積神經網路為基礎的模組（SeisPolor），用於分類地震連續資料中的 P 波極性，進而用 FPFIT 等方法自動解析地震震源機制。而在強化部分著重將波相到時偵測模型以 PyTorch 改寫以提升系統靈活性，提高模型調整與實驗的開發效率。 研究結果顯示，在自動識別P波極性的模型表現優異，精確率達 95%。將預測極性解析成震源機制，並考量測站方位角與距離的包覆度門檻後，採用 Kagan 測試方法評估預測與真實震源機制解之間的相似性，Kagan 測試方法以最小旋轉角度小於40度為預測正確的標準，震源機制解的結果顯示近 80% 的準確率。此外，在重新設計與訓練的波相到時偵測模型，在 P 波和 S 波都解決提早挑選的現象，並使精確率提升約 8%。 此外，本研究以資料管線為主軸重新設計整體系統的自動化流程，系統在實現過程中廣泛借鑒了資訊工程的先進技術，整合了硬體、系統環境、資料庫、資料管線、模型開發、任務監控以及資料可視化、Web UI 互動等多方面技術。 本研究在原有基礎上達到系統擴展、強化和整合。目前已實現了從地震連續資料處理到挑波、定位、規模計算及震源機制解析的半自動化流程。並經實驗與調整後，大幅提升波相到時偵測模型的精確率。結合軟體技術的優勢，本系統不僅加快了資料處理速度，系統重構後強調的管線化設計，也為未來提供了快速開發的可能性，使模型性能不斷提升。對於地震活躍的台灣，該系統將大幅提升地震觀測的速度與品質，抓回大量小規模地震資訊，以高時空解析度的地震目錄協助評估活動構造。

To address the significant challenges posed by the rapid increase in dense seismic networks and seismic data, this study focuses on applying deep learning techniques to enhance the automation and efficiency of seismic data processing, particularly advancing microseismic monitoring developments（0.8M+）. Since 2018, the Structural Seismology Lab（SGYLAB） at National Taiwan University has developed the SeisBlue system, which has been successfully applied in multiple cases since 2020 to achieve regional near-real-time earthquake localization. This research is dedicated to developing a more refined third-generation system, with an emphasis on system expansion, enhancement, and integration. A key development is a CNN-based module（SeisPolor）, designed for classifying P-wave polarity in continuous seismic data, subsequently utilizing the FPFIT method to automatically analyze earthquake focal mechanisms. The enhancement efforts include redesign the phase picking model in PyTorch to improve system flexibility and increase the efficiency of model adjustments and experimental development. The results demonstrate that the model for automatic identification of P-wave polarity performs excellently, with a precision of 97%. By converting the predicted polarities into focal mechanisms and considering coverage thresholds based on station azimuth and distance, the Kagan test method is employed to evaluate the similarity between the predicted and actual focal mechanisms. This method uses a minimum rotation angle of less than 40 degrees as the criterion for a correct prediction, showing that nearly 80% of the focal mechanisms are accurate. Moreover, the redesigned and retrained picking model solved the issue of early picking and showed an 8% increase in precision for both P-wave and S-wave. Furthermore, the study redesigned the overall system automation workflow around a data pipeline, extensively drawing on advanced information engineering technologies. The integration encompasses hardware, system environment, databases, data pipelines, model development, task monitoring, data visualization, and Web UI interaction. Building on the existing foundation, this study achieves system expansion, enhancement, and integration. A semi-automated workflow from continuous seismic data processing to picking, localization, magnitude estimation, and focal mechanism analysis has been realized. Through experiments and adjustments, the picking model's accuracy has been significantly improved. Leveraging software technology advantages, the system not only speeds up data processing but also emphasizes a pipeline design in its reconstruction, providing potential for rapid development and continual performance enhancement. For Taiwan, a region prone to seismic activity, this system can significantly enhance the speed and quality of earthquake monitoring, capturing a substantial amount of small earthquake data. It will assist in the evaluation of active structures through a high spatio-temporal resolution earthquake catalog.