航空業處於極度複雜且風險極大的營運環境中，因此航空風險評估變得至關重要。未經充分評估的航班可能導致重大事故的發生。飛航操作風險評估系統（Flight Operations Risk Assessment System, FORAS）透過計算各種風險相關數據，提供了對航班在起飛和降落兩個時段的風險值的評估。這些風險值反映了航空事故發生的機率，風險值越高代表航班事故發生的機率越大。該系統評估後的風險值為飛行員和飛航安全單位制定操作計劃的參考。
    然而，FORAS系統對於針對風險較高的因素所提供的避險措施較為薄弱。因此，本研究使用生成式AI（ChatGPT）分析FORAS的決策風險值數據，以即時提供降低風險的建議。為了讓ChatGPT能夠更好地完善FORAS系統，需要使ChatGPT了解FORAS整個系統，使其能夠自主判斷風險值的高低並提供相應的避險建議。此外，本研究探討將生成式人工智慧（ChatGPT）應用於航空風險分析的可行性，這種智能化的風險評估模式有望提高整體的判斷準確性和效能，同時提供更及時的風險警示。
    在新的應用場景下，生成式人工智慧的技術可以自動判斷FORAS系統生成的風險值，並提供風險值較高所應對的避險建議給飛行員參考。這項技術的引入有助於提高航空安全水平，降低事故風險。

    The aviation industry operates in an extremely complex and high-risk environment, making aviation risk assessment critically important. Flights that are not adequately assessed can lead to major accidents. The Flight Operations Risk Assessment System (FORAS) calculates various risk-related data to provide an evaluation of the risk values during takeoff and landing phases of flights. These risk values reflect the probability of aviation accidents, with higher values indicating a greater likelihood of flight accidents. The assessed risk values from this system serve as references for pilots and aviation safety units in devising operational plans.
    However, the FORAS system provides relatively weak mitigation measures for factors with higher risks. Therefore, this study employs Generative AI (ChatGPT) to analyze the decision risk data of FORAS, offering real-time suggestions for risk reduction. To enable ChatGPT to better enhance the FORAS system, it is essential to familiarize ChatGPT with the entire FORAS system, enabling it to autonomously assess the level of risk and provide corresponding mitigation recommendations. Furthermore, this study explores the feasibility of applying Generative Artificial Intelligence (ChatGPT) to aviation risk analysis. This intelligent risk assessment model holds the potential to enhance overall judgment accuracy and efficiency while delivering more timely risk alerts.
    In new application scenarios, the technology of Generative Artificial Intelligence can automatically determine the risk values generated by the FORAS system and provide pilots with mitigation recommendations tailored to higher-risk scenarios. The introduction of this technology contributes to enhancing aviation safety levels and reducing the risk of accidents.

目錄
第一章 緒論	1
第二章 文獻回顧	3
2-1 AI技術在安全風險評估的應用	3
2-2生成式AI的架構	4
2-3航空安全風險評估之文獻	8
2-4航空事故肇因之分析	9
2-4-1 錯誤鏈法則 (error chain)	10
2-4-2 多米諾骨牌效應(Domino Effect)	12
2-4-3 乳酪理論(Swiss Cheese Model or Reason’s Model)	13
2-4-4 能量釋放理論(energy elease theory)	14
2-4-5 風險管理天秤理論(scales theory)	14
2-5 FORAS及其他飛航風險評估之相關研究	15
第三章 研究方法	17
3-1 飛行操作風險評估系統(FORAS)	17
3-2 ChatGPT	20
3-2-1 Transformer模型	21
3-2-2預訓練(Pre-training)模型	22
3-2-3 微調（Fine-tuning）	23
3-2-4強化學習（Reinforcement Learning）	23
3-3 資料前處理	24
3-4 系統實作	25
第四章 研究分析與結果	27
4-1 資料集	27
4-2 實驗結果	29
4-2-1 模型訓練	29
4-2-2 ChatGPT問答結果與分析	30
4-2-3 研究發現	38
第五章 結論與建議	41
參考資料	42
圖目錄
圖 1 航空事故與航空安全因素關係流通示意圖	1
圖 2 連鎖型航空事故的模型	9
圖 3 匯集型航空事故的模型	9
圖 4 複合型航空事故的模型	10
圖 5 航空事故錯誤鏈法則	10
圖 6 航空事件與事故比例	11
圖 7 航空事故骨牌效應	12
圖 8 乳酪理論	13
圖 9 民用航空運輸業重大飛航事故分類統計	17
圖 10 風險模型示意圖	18
圖 11 GPT架構	20
圖 12 Transfomer 架構	21
圖 13 Masked Self-Attention示意圖	22
圖 14 ChatGPT FineTune 頁面	25
圖 15 2000筆資料集的模型訓練結果	27
圖 16 模型訓練結果	29
圖 17 47筆資料集的模型訓練結果	38
圖 18 37筆資料集的模型訓練結果	38
圖 19 27筆資料集的模型訓練結果	39
圖 20 200筆資料集的模型訓練結果	40
表目錄
表 1 生成式AI產品	4
表 2 生成模型的性能比較	6
表 3 國內文獻有關航空風險課題	8
表 4 航空安全風險評估系統各個優缺點	16
表 5  C2 Rule Table	19
表 6 「What is FORAS?」模型回答對照表	31
表 7 「What is the structure of FORAS?」模型回答對照表	31
表 8 「What are the main categories of risk factors in FORAS?」模型回答對照表	31
表 9 「How are risk values defined in FORAS?」模型回答對照表	32
表 10「What is the criterion for high risk values in FORAS?」模型回答對照表	32
表 11「How is the tree structure of FORAS generated?」模型回答對照表	33
表 12「What are the sub-risks of crew functionality?」模型回答對照表	33
表 13「What is the upper-level risk of intercrew communication?」模型回答對照表	33
表 14 數據判定風險值的模型對照表 	35
表 15 不同資料集的模型回答結果	39
表 16 不同資料集的指令對照表	40

參考資料
[1] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever,” Improving Language Understanding by Generative Pre-Training”(2018)
[2] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, Percy Liang,” Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution”(2022)
[3] Ashish Vaswan, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser.”Attention Is All You Need”(2017)
[4] F. Gonçalves, B. Ribeiro, O. Gama, A. Santos, A. Costa, B. Dias, J. Macedo, and M. J. Nicolau, “A systematic review on intelligent intrusion detection systems for vanets,” in 2019 11th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), 2019, pp. 1–10. DOI: 10.1109/ICUMT48472.2019. 8970942.
[5] Ghassemi, M., Naumann, T., Doshi-Velez, F., Brimmer, N., Joshi, R., Rumshisky, A., & Szolovits, P. (2014, August). Unfolding Physiological State: Mortality Modelling in Intensive Care Units. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining , 75-84, ACM.
[6] G. Mohi-ud-din, Nsl-kdd, 2018. DOI: 10.21227/425a-3e55. [Online]. Available: https: //dx.doi.org/10.21227/425a-3e55.
[7] Goodfellow, Ian J.; Pouget-Abadie, Jean; Mirza, Mehdi; Xu, Bing; Warde-Farley, David; Ozair, Sherjil; Courville, Aaron; Bengio, Yoshua.(2014).”Generative Adversarial Networks.”
[8] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, Surya Ganguli.(2015,Nov).” Deep Unsupervised Learning using Nonequilibrium Thermodynamics”
[9] Jay Alammar,”The Illustrated GPT-2 (Visualizing Transformer Language Models)”,
網址:http://jalammar.github.io/illustrated-gpt2/，上網日期：2023 /4 /15
[10] Jensen, F.V., Introduction to Bayesian Networks, London, UK: University-College London Press (1995).
[11] J. Song, H. Takakura, Y. Okabe, M. Eto, D. Inoue, and K. Nakao, “Statistical analysis of honeypot data and building of kyoto 2006+ dataset for nids evaluation,” in Proceedings of the First Workshop on Building Analysis Datasets and Gathering Experience Returns for Security, 2011, pp. 29–36, ISBN: 9781450307680. DOI: 10.1145/1978672.1978676.
[12] Kingma D P, Welling M. Auto-Encoding Variational Bayes[J]. stat, 2014, 1050: 10.
[13] Laurent Dinh, David Krueger, Yoshua Bengio (2014).”NICE: Non-linear Independent Components Estimation”
[14] Luxhøj, J.T., “Probabilistic Causal Analysis for System Safety Risk Assessments in Commercial Air Transport”, in Proceedings of the Workshop on Investigating and Reporting of Incidents and Accidents (IRIA), Williamsburg , VA, Sep 2003.
[15] Manela, A., & Moreira, A. (2017). “News implied volatility and disaster concerns.” Journal of Financial Economics, 123(1), 137-162.
[16] Michael, H., Osborne, D.M., Ross, D., Boyd, D., and Brown, B.G., “The Flight Operations Risk Assessment System”, Proceedings of the SAE Advances Safety Conference. Society of Automotive Engineers, 1999.
[17] Micheal, H., John, M., “Flight Operations Risk Assessment System (FORAS)”,International Air Safety Seminar, Flight Safety Foundation, 2002
[18] Sims, C. A. (1980), Comparison of Interwar and Postwar Business Cycles: Monetarism Reconsidered, American Economic Review 70 (2), 250–257
[19] Takagi, T. and Sugeno, M. “Fuzzy identication of systems and its application to modeling and control”, IEEE Transactions on Systems, Man, and Cybernetics, 15, pp. 116- 132 (1985)
[20] U.S. General Accounting Office “Aviation safety: U.S. efforts to implement flight operational quality assurance programs", Flight Safety Digest, 17, pp. 1-47 (1998).
[21] uxhoj, J.T. and Coit, D.W. “Modeling low probability/high consequence events: an aviation safety risk model", In Proceedings of the Annual Reliability and Maintainability Symposium, Newport Beach, CA, Jan. 23-26, pp. 215-221 (2006).
[22] Wiegmann, D.A., Shappell, S.A., “The Human Factors Analysis and Classification System (HFACS)”, U.S. Department of Transportation, Office of Aviation Medicine, Washington, DC (2000).
[23] Y. Shen, H. Yun, Z.C. Lipton, Y. Kronrod, A. Anandkumar, Deep active learning for named entity recognition, pp. 1-15, 2017.
[24] 能力雜誌，<員工到領導者，從AI到生成式AI的思維轉變>，網址: https://mymkc.com/article/content/25133，上網日期:2024/01/05
[25] 仁靜怡，航空公司飛安管理之研究，民航季刊，第一卷第一期，頁73-92，民國88年3月
[26] 林長慶，風險管理-中正國際航空站，交通部民用航空局擴大行務主管座談會簡報資料，民國90年4月
[27] 張有恆、李文魁，航空安全風險評估模式之建構，運輸計畫季刊，第三十四卷第一期，頁145-176，民國94年3月。
[28] 張有恆、李文魁，模糊FMECA應用在飛安風險評估模式之研究，民航季刊，第七卷第一期，頁1-22，民國94年3月
[29] 陸鵬舉、嵇允禪，國際航空器飛安事故模型建立及預測之研究，國立成功大學航空太空研究所，民國85年3月
[30] 陳奕翔，民96，『飛行操作風險評估系統(FORAS)之風險因素權重分析』，國立成功大學民航研究所碩士論文。
[31] 張有恆，「現代軌道運輸：放眼世界的軌道系統與管理」，初版，台北 縣，人人圖書出版，第250頁~第259頁，民國91年。
[32] 國家運輸安全調查委員會，台灣飛安統計2012-2021，網址: https://www.ttsb.gov.tw/media/5636/%E5%8F%B0%E7%81%A3%E9%A3%9B%E5%AE%89%E7%B5%B1%E8%A8%882012-2021.pdf