臺灣博碩士論文加值系統

本研究以RNN建立用於預測及判別電離層地震前兆之類神經網路模型。利用電離層全電子含量(TEC)具有時間序列的特性，將數據轉換成RNN之訓練資料，建立類神經網路模型。並先探討影響模型判別結果之參數，分別為訓練次數、批次大小、時間序列長度和訓練資料長度四項，以2017/10/6之地震測試各項參數較適用範圍。判別標準為測試資料大於訓練資料平均值上下三倍標準差者為異常值，比較正常日(地震發生前16-30天)和異常日(地震發生前15天)異常值個數。接著實際測試四組於2016至2017年間地震規模大於6之地震前的TEC資料，且從訓練資料到預測資料期間皆沒有發生規模6以上之地震。在訓練次數為200次、批次大小為100、時間序列長度為2天、訓練資料長度45天的參數下，四組地震案例皆符合在異常日的異常值個數多於正常日。比較前人使用之判定方法(中位數法)和本研究之RNN模型，本研究之RNN模型對於判別是否有發生規模6以上之地震，皆有較好的判別能力。

This study uses Recurrent Neural Network (RNN) to establish a neural network model for analysis ionospheric data to identify earthquake precursors. Using the time series characteristic of ionospheric electron content (TEC), the neural network model is built through converting the data to RNN training data. Among those, the parameters of model such as training times (epochs), batch size, time series length, and the length of training data are discussed first. These parameters range are individually tested on the earthquake of 2017/10/6. The abnormal value is defined as three times the standard deviation of the average value of the training data. In this work, the abnormal value is employed to compare the number of abnormal values on the normal day (16-30 days before the earthquake) and the abnormal day (15 days before the earthquake). Then, four real TEC data sets in 2016 and 2017 are conducted as training and forecast data to identify a particular 6 magnitude earthquake following the data, and there are no any earthquake above magnitude of 6 and above in the data sets. The parameters of RNN are set as followings: training epoch is 200, batch size is 100, time series duration is 2 days, and training data length is 45 days. The results of identification for the four data sets meet the trend that abnormal values on the abnormal days are more than normal days. In comparison to previous application of median method and our study of RNN model, the RNN model prove to have a better distinguishing ability for earthquake with a magnitude of 6 or above.

摘要 i
Abstract ii
誌謝 iii
目錄 v
表目錄 ix
圖目錄 x
第一章　緒論 1
1.1 前言 1
1.2 研究動機 2
1.3 研究目的 3
1.4 研究架構 3
第二章 文獻回顧 5
2.1 電離層 5
2.2 電離層與地震關係 6
2.3 電離層預測地震相關方法 7
2.4 類神經網路的基本介紹 8
2.4.1 生物神經網路 9
2.4.2 類神經網路 9
2.5 遞迴式類神經網路之應用 13
第三章　研究方法 14
3.1 類神經網路 14
3.1.1 神經元模型 14
3.1.2 遞迴類神經網路(RNN) 16
3.1.3 長短期記憶 (LSTM) 19
3.2 TensorFlow和Keras 23
3.3 時間序列模型用於TEC預測 25
3.3.1 TEC時序性變化 26
3.3.2 TEC時間序列模型推導 29
3.3.3 模型預測方式 31
第四章　參數探討與案例分析 34
4.1 模型建立與資料預處理 34
4.1.1 建立類神經網路模型 34
4.1.1 資料預處理 38
4.2 參數探討 38
4.2.1 訓練次數 39
4.2.2 批次大小 40
4.2.3 時間序列長度 41
4.2.4 訓練資料長度 44
4.3 預測結果與討論 46
4.3.1 案例一(地震日2017/4/30) 46
4.3.2 案例二(地震日2016/10/6) 47
4.3.3 案例三(地震日2016/5/12) 48
4.3.4 案例四(地震日2016/2/6) 49
4.3.5 預測結果之問題與討論 50
4.4 中位數法 51
4.5 結果比較 53
第五章 結論與未來展望 55
5.1 結論 55
5.2 未來展望 56
參考文獻 57
附錄A-案例一(2017/4/30)之預測圖 60
A.1 參數組(訓練資料45天、時間序列1天) 61
A.2 參數組(訓練資料45天、時間序列2天) 65
A.3 參數組(訓練資料45天、時間序列3天) 69
A.4 參數組(訓練資料45天、時間序列7天) 73
A.5 參數組(訓練資料60天、時間序列1天) 77
A.6 參數組(訓練資料60天、時間序列2天) 81
A.7 參數組(訓練資料60天、時間序列3天) 85
A.8 參數組(訓練資料60天、時間序列7天) 89
附錄B-案例二(2016/10/6)之預測圖 93
B.1 參數組(訓練資料30天) 94
B.2 參數組(訓練資料45天、時間序列1天) 98
B.3 參數組(訓練資料45天、時間序列2天) 102
B.4 參數組(訓練資料45天、時間序列3天) 115
B.5 參數組(訓練資料45天、時間序列7天) 119
B.6 參數組(訓練資料60天、時間序列1天) 123
B.7 參數組(訓練資料60天、時間序列2天) 127
B.8 參數組(訓練資料60天、時間序列3天) 131
B.9 參數組(訓練資料60天、時間序列7天) 135
附錄C-案例三(2016/5/12)之預測圖 139
C.1 參數組(訓練資料45天、時間序列1天) 140
C.2 參數組(訓練資料45天、時間序列2天) 144
C.3 參數組(訓練資料45天、時間序列3天) 148
C.4 參數組(訓練資料45天、時間序列7天) 152
C.5 參數組(訓練資料60天、時間序列1天) 156
C.6 參數組(訓練資料60天、時間序列2天) 160
C.7 參數組(訓練資料60天、時間序列3天) 164
C.8 參數組(訓練資料60天、時間序列7天) 168
附錄D-案例四(2016/2/6)之預測圖 172
D.1 參數組(訓練資料45天、時間序列1天) 173
D.2 參數組(訓練資料45天、時間序列2天) 177
D.3 參數組(訓練資料45天、時間序列3天) 181
D.4 參數組(訓練資料45天、時間序列7天) 185
附錄E 189

[1] Derr, J.S., Earthquake lights: a review of observations and present theories. Bull. seis. Soc. Am, 1973. 63: p. 2177-2187.
[2] Derr, J.S., Luminous phenomena and their relationship to rock fracture. Nature, 1986. 321: p. 470-471.
[3] Heraud, J.A.L.A., Co-seismic luminescence in Lima, 150 km from the epicenter of the Pisco, Peru earthquake of 15 August 2007. Nat. Hazards Earth Syst. Sci., 2011. 11: p. 1025-1036.
[4] Finkelstein, D.P., J., Earthquake Ligthning. Nature, 1970. 228: p. 759-760.
[5] Lockner, D.A.J., M. J. S.; Byerlee, J. D., A mechanism to explain the generation of earthquake lights,. Nature, 1983. 302: p. 28-33.
[6] Freund, F., Time-resolved study of charge generation and propagation in igneous rocks. J. Geophys. Res.,, 2000. 105: p. 11001-11019.
[7] Pulinets, S.A.B., K. A., Ionospheric precursors of earthquake. 2004.
[8] St-Laurent, F.D., J. S.; Freund, F. T., Earthquake lights and the stress-activation of positive hole charge carriers in rocks. Phys. Chem. Earth, 2006. 31: p. 305-312.
[9] Davies, K., Ionospheric Radio,. 1990: Peregrinus Ltd.
[10] Kelley, M.C., The Earth’s Ionosphere. 1989.
[11] Gokhberg, M.B.G., I. L.; Rozhnoy, A. A.; Marenko, V. F.; Yampolsky, V. S.; Ponomarev, E. A., Study of seismic influence on the ionosphere by super long-wave probing of the Earth-ionosphere waveguide. Phys. Earth Planet. Inter, 1989. 57: p. 64-67.
[12] Molchanov, O.A.H., M. , Subionospheric VLF signal perturbations possibly related to earthquakes,. J. Geophys. Res., 1998. 103: p. 17489-17504.
[13] Fujiwara, H.K., M.; Ikeda, M.; Liu, J. Y.; Sakata, H.; Chen, Y. I.; Ofuruton, H.; Muramatsu, S.; Chuo, Y. J.; Ohtsuki, Y. H. , Atmospheric anomalies observed during earthquake occurrences,. Geophys. Res. Lett, 2004. 31: p. L17110.
[14] Shvets, A.V.H., M.; Molchanov, O. A.; Ando, Y., A study of ionospheric response to regional seismic activity by VLF radio sounding. Phys. Chem. Earth,, 2004. 29: p. 627-637.
[15] Liu, J.Y., et al., Variations of ionospheric total electron content during the Chi‐Chi Earthquake. Geophys. Res.Lett., 2001. 28: p. 1383–1386,.
[16] Liu, J.Y.C., Y. I.; Chuo, Y. J.; Chen, C. S., A statistical investigation of preearthquake ionospheric anomaly. J. Geophys. Res., 2006. 111: p. 1–5,.
[17] Liu, J.Y., Y. I. Chen, C. H. Chen, C. Y. Liu, C. Y. Chen, M. Nishihashi, J. Z. Li, Y.Q. Xia, K.-I. Oyama, K. Hattori, C. H. Lin, , Seismo-ionospheric GPS TEC anomalies observed before the 12 May 2008 Mw7.9 Wenchuan earthquake. J.Geophys. Res., 2009. 114: p. A04320.
[18] Molchanov, O.A., et al., Precursory effects in the subionospheric VLF signals for the Kobe earthquake. Physics of the Earth and Planetary Interiors, 1998. 105(3): p. 239-248.
[19] Sasmal, S., S.K. Chakrabarti, and S. Ray, Unusual behavior of Very Low Frequency signal during the earthquake at Honshu/Japan on 11 March, 2011. Indian Journal of Physics, 2014. 88(10): p. 1013-1019.
[20] Kuo, C.L., et al., Ionosphere plasma bubbles and density variations induced by pre-earthquake rock currents and associated surface charges. Journal of Geophysical Research: Space Physics, 2011. 116(A10).
[21] Mei, L. and P. Michel, Statistical analysis of an ionospheric parameter as a base for earthquake prediction. Journal of Geophysical Research: Space Physics, 2013. 118(6): p. 3731-3739.
[22] McCullch, W.S. and W. Pitts, A logical calculus of the ideas immanent in nervous activity. The bulletin of mathematical biophysics, 1943. 5(4): p. 115-133.
[23] Hebb, D.O., The Organization of Behavior. 1949, New York: Wiley & Sons.
[24] Rosenblatt, F., The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 1958. 65(6): p. 386-408.
[25] Haykin, S., Neural Network：A Comprehensive Fundation. 1999, New York: Prentice Hall Intl.
[26] Cruse, H., Neural Networks as Cybernetic Systems. 1996: Thieme Medical Publishers, Incorporated.
[27] 鄭允中, 基於長短期記憶遞迴類神經網路之新台幣兌美元匯率預測模型, in 資訊工程學系. 2017, 台灣大學.
[28] 張斐章；黃浩倫；張麗秋, 回饋式類神經網路於河川流量推估之應用. 農業工程學報, 2001. 47: p. 32-39.
[29] 張國興；李亞東；張磊；樊慶富；李想, 基於SDZ-RNN的出租車行目的地預測方法. 計算機工程與應用, 2018. 54.
[30] 張杰 and 張艷霞, 基於CAPSO-RNN的光伏系統短期發電量預測. 中國電力, 2012. 45: p. 87-91.
[31] 笱程成, et al., 一種基於RNN的社交消息爆發預測模型. 軟體學報, 2017. 28: p. 3030-3042.
[32] Jang, J.S.R., C.T. Sun, and E. Mizutani, Neuro-Fuzzy and Soft Computing-Acomputational Approach to Learning and Mechine Tntelligence. 1997: Prentice hall.
[33] 張斐章, 張麗秋, and 黃浩倫, 類神經網路. 2003: 東華書局.
[34] Hochreiter, S. and J. Schmidhuber, Long short-term memory. Neural Computation, 1997. 9(8): p. 1735-1780.
[35] Bengio, Y., P. Simard, and P. Frasoni, Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, 1994. 5(2): p. 157-166.
[36] Hochreiter, S., et al., Gradient flow in recurrent nets: the difficulty of learning long-term dependencies. 2001.
[37] Pascanu, R., T. Mikolov, and B. Y. On the difficulty of training recurrent neural networks. in International Conference on Machine Learning. 2013.
[38] 林大貴, TensorFlow+Keras深度學習人工智慧實務應用. 2017: 博碩.
[39] Komjathy, A., Global Ionospheric Total Electron Content Mapping Using the Global Positioning System, in Department of Geodesy and Geomatics Engineering Technical Report 1997: Canada. .
[40] K., J.H., et al., Ionospheric tidal waves observed from Global Ionosphere Maps: analysis of total electron content. Journal of Geophysical Research: Space Physics. 0(ja).
[41] Liu, J.Y., Y.I. Chen, S.A., Pulinets, Y.B., Tsai, Y.J., Chuo,, Seismoionospheric signatures prior to M ≧ 6.0 Taiwan earthquakes. Geophys. Res. Lett., 2000. 27: p. 3113-3116.
[42] 林鈺翔, 利用全球定位系統觀測電離層地震前兆. 2004, 國立中央大學.
[43] 蘇泳誌, 電離層地震前兆與震源機制. 2009, 國立中央大學.