現地型強震預警系統主要目的在於提前發出精準的警報，讓近震央區域有時間做好預防措施，以減少強震波帶來的傷亡與財損。目前現地型強震預警系統需要先判斷P波到時，之後利用P波到時後之訊號預測PGA，據以對單一測站發出警報。目前判斷P波到時之模型以及預測PGA之模型係互相獨立運作。本研究為開發一模型可以同時處理這兩個問題，並設定在即時系統上運行。本研究以Liao等人所開發之即時標記地震波相模型架構為基礎，開發一可以同時標記地震波相並進行單站法強震預警之模型，並探討在不同輸入地表加速度歷時尺度組合之效果，所考慮的組合包括Z-score、Z-score&0.8gal、Z-score&2.5gal、Z-score&8gal、Z-score&25gal、Z-score&80gal、Z-score&250gal、Z-score&800gal和Z-score&原始加速度歷時。 評估模型效果有四種方式， (一)、評估九個模型對P波到時之準確率、召回率和F1-score，判斷訊號是否為地震訊號之準確率、召回率和F1-score，預測PGA是否大於25gal之準確率、召回率和F1-score，並觀察每個指標之九個模型在P波到時後0.1到3秒間百分比之變化。(二)、觀察九個模型在P波到時後0.1到3秒間對真實PGA在0-8gal和8-25gal之誤報率、真實PGA在25-80gal、80-250gal、250-800gal和高於800gal之漏報率。(三)、綜合考慮地震波形檢測和單站預警之結果是否能降低模型對noise誤報或是對地震訊號漏報。(四)、九個模型對P波到時後0.1到3秒間綜合發報之準確率、召回率和F1-score。最後，本研究選擇Z-score+正負2.5gal並用P波到時後2秒之訊號發報。Z-score&2.5gal預測池上地震觸發測站在無容許誤差下，P波到時後2秒到3秒之整體表現之準確率為89.3%，召回率為38.5%。容許誤差下，P波到時後2秒到3秒之整體表現之準確率為98.7%，召回率為93.7%。

The primary objective of an on-site early warning system for strong earthquakes is to issue accurate alerts in advance, allowing the regions near the epicenter to take precautionary measures to minimize casualties and property damage caused by strong seismic waves. Currently, the on-site early warning system involves two independent processes: determining the arrival time of P-waves and predicting Peak Ground Acceleration (PGA) based on the P-wave arrival time. This study aims to develop a model that can simultaneously address both of these issues and operate in real-time.Built upon the real-time seismic phase identification framework developed by Liao et al., this research establishes a model that can concurrently identify seismic wave phases and perform single-station early earthquake warning. It explores the effectiveness of various combinations of input ground acceleration time series scales, including Z-score, Z-score & 0.8 gal, Z-score & 2.5 gal, Z-score & 8 gal, Z-score & 25 gal, Z-score & 80 gal, Z-score & 250 gal, Z-score & 800 gal, and Z-score & raw acceleration time series.The model's effectiveness is evaluated using four approaches: (1) assessing the accuracy, recall, and F1-score of the nine models for P-wave arrival time, the accuracy, recall, and F1-score of identifying seismic signals, and the accuracy, recall, and F1-score of predicting PGA exceeding 25 gal, observing the percentage variation of the nine models for each indicator between 0.1 and 3 seconds after P-wave arrival time; (2) observing the false positive rates for real PGAs in the ranges of 0-8 gal and 8-25 gal and the false negative rates for real PGAs in the ranges of 25-80 gal, 80-250 gal, 250-800 gal, and above 800 gal between 0.1 and 3 seconds after P-wave arrival time; (3) comprehensively considering whether the combination of seismic waveform detection and single-station warning can reduce noise false alarms or seismic signal missed alarms; and (4) the combined accuracy, recall, and F1-score for issuing alerts between 0.1 and 3 seconds after P-wave arrival time.Ultimately, this study selects Z-score + ±2.5 gal and issues alerts using signals 2 seconds after the P-wave arrival time. Under the Z-score & 2.5 gal prediction, the overall accuracy of the seismic trigger station's performance between 2 and 3 seconds after P-wave arrival time is 89.3%, with a recall rate of 38.5%. With allowable error, the overall accuracy is 98.7%, with a recall rate of 93.7%.

[1] W.-Y. Liao, E.-J. Lee, D.-Y. Chen, P. Chen, D. Mu, and Y.-M. Wu, "RED-PAN: Real-Time Earthquake Detection and Phase-Picking With Multitask Attention Network," IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-11, 2022, doi: 10.1109/tgrs.2022.3205558.
[2] L. Geiger, "Probability method for determination of earthquake epicenters from arrival time only," 1912.
[3] D.-Y. Chen, T.-L. Lin, H.-C. Hsu, Y.-C. Hsu, and N.-C. Hsiao, "An approach to improve the performance of the earthquake early warning system for the 2018 Hualien earthquake in Taiwan," Terr. Atmos. Ocean. Sci, vol. 30, pp. 423-433, 2019.
[4] R. M. Allen, P. Gasparini, O. Kamigaichi, and M. Bose, "The status of earthquake early warning around the world: An introductory overview," Seismological research letters, vol. 80, no. 5, pp. 682-693, 2009.
[5] M. Böse, T. Heaton, and E. Hauksson, "Rapid estimation of earthquake source and ground‐motion parameters for earthquake early warning using data from a single three‐component broadband or strong‐motion sensor," Bulletin of the Seismological Society of America, vol. 102, no. 2, pp. 738-750, 2012.
[6] Y.-M. Wu and H. Kanamori, "Rapid assessment of damage potential of earthquakes in Taiwan from the beginning of P waves," Bulletin of the Seismological Society of America, vol. 95, no. 3, pp. 1181-1185, 2005.
[7] T.-Y. Hsu et al., "Rapid on-site peak ground acceleration estimation based on support vector regression and P-wave features in Taiwan," Soil Dynamics and Earthquake Engineering, vol. 49, pp. 210-217, 2013.
[8] 黃昭文, "卷積神經網路應用於初達波預估當站最大地表加速度之研究," 碩士, 營建工程系, 國立臺灣科技大學, 台北市, 2020.
[9] H. Robbins, & Monro, S, "A stochastic approximation method," The annals of mathematical statistics, 1951.
[10] L. Bottou, "Large-scale machine learning with stochastic gradient descent," In Proceedings of COMPSTAT'2010, pp. 177-186, 2010.
[11] J. Duchi, Hazan, E., & Singer, Y, "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization," Journal of Machine Learning Research, 2011.
[12] D. P. Kingma, & Ba, J., "ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION," International Conference on Learning Representations 2015.
[13] L. Liu, "On the Variance of the Adaptive Learning Rate and Beyond," arXiv:1908.03265, 2021.
[14] W. I. Chigozie Enyinna Nwankpa, Anthony Gachagan, and Stephen Marshall, "Activation Functions: Comparison of Trends in Practice and Research for Deep Learning," arXiv, 2018.
[15] S. Ruder, "An Overview of Multi-TaskLearning in Deep Neural Networks," arXiv:1706.05098, 2017.
[16] M. Crawshaw, "Multi-task Learning With Deep Neural Networks A Surve," arXiv:2009.09796, 2020.
[17] S. Liu, "End-to-End Multi-Task Learning with Attention," arXiv:1803.10704, 2019.
[18] S. M. Mousavi, Y. Sheng, W. Zhu, and G. C. Beroza, "STanford EArthquake Dataset (STEAD): A Global Data Set of Seismic Signals for AI," IEEE Access, vol. 7, pp. 179464-179476, 2019, doi: 10.1109/access.2019.2947848.
[19] W.-Y. Liao, E.-J. Lee, D. Mu, P. Chen, and R.-J. Rau, "ARRU Phase Picker: Attention Recurrent-Residual U-Net for Picking SeismicP- andS-Phase Arrivals," Seismological Research Letters, vol. 92, no. 4, pp. 2410-2428, 2021, doi: 10.1785/0220200382.
[20] W.-Y. Liao, E.-J. Lee, D. Mu, and P. Chen, "Toward Fully Autonomous Seismic Networks: Backprojecting Deep Learning-Based Phase Time Functions for Earthquake Monitoring on Continuous Recordings," Seismological Research Letters, vol. 93, no. 3, pp. 1880-1894, 2022, doi: 10.1785/0220210274.