應用序列到序列方法建立航空器重落地及長平飄之肇因分類模型

Establishing a Causal Analysis Model of Hard Landing and Long Landing Events by Using Sequence to Sequence Method

童奕城(Yi-Cheng Tong)

指導教授 ： 蕭育霖

中原大學/電機資訊學院/工業與系統工程研究所/碩士(2021年)

https://doi.org/10.6840/cycu202100010

摘要

航空安全是航空公司最重視的環節之一，在此前提下，民航業者普遍使用快速存取記錄器(QAR)蒐集飛航數據，再根據QAR資料建立飛航作業品質保證系統(FOQA)。研究航機飛行時的各項參數是否超限，並對航機操作行為進行監測及分析，以提高飛行安全及整體營運效率。然而，目前針對FOQA參數超限事件的肇因判斷主要仰賴專家進行分析，不僅需花費許多時間與人力，也無法保證其結果的客觀性。因此，本研究與國內民航業者合作，從其歷年參數超限事件中選取落地階段的重落地及長平飄事件作為研究課題，收集兩事件之QAR數據資料與事後分析報告作為原始研究資料。本研究依據過去相關研究之專家訪談結果決定重要影響參數以及可能肇因分類後，運用類神經網路中的序列到序列(Seq2seq)方法搭配循環神經網路(RNN)、長短期記憶(LSTM)、雙向長短期記憶(BiLSTM)與GRU(Gated Recurrent Unit)四種類神經網路方法，建立Force與Timing兩類常發肇因之判斷模型。研究結果顯示，在Force模型方面，Seq2seq搭配BiLSTM可使肇因判斷之準確率達到79%，精確率、召回率與F1-Score等指標皆達到0.80以上。Timing模型方面，Seq2seq搭配GRU之準確率則可達到82%，精確率、召回率與F1-Score等指標皆達到0.83以上。與其他相關肇因判斷分析文獻相比，本研究的準確度更高，表示使用Seq2seq方法可提升肇因判斷結果的準確率及可信度。建議未來可持續使用機器學習方法，例如注意力機制，建立更完善的飛航操作肇因自動判讀機制，應能有效提升航空公司的分析效率，即時對飛安事件做出肇因分析，促進並落實飛航安全管理。

關鍵字

飛航作業品質保證系統 ； 重落地 ； 長平飄 ； 類神經網路 ； 序列到序列

並列摘要

Flight safety is the most important aspect for any airlines. To assure the safety of flight operation, civil airlines commonly use onboard Quick Access Recorders (QARs) to collect flight data, and develop Flight Operations Quality Assurance (FOQA) system to monitor the daily flight conditions. The purpose of FOQA is to assure flight safety and to improve overall operational efficiency by detecting the exceedance of the flight parameters during daily flight. However, the determination of the operational causes of the parameter exceedance events relies majorly on subject matter experts such as senior pilots. The analysis is costly, time consuming and sometimes lack of objectiveness. The purpose of this study was to establish a causal classification model to replace the manual analysis. Two kinds of parameter exceedance events, i.e. hard landing and long landing, were selected in the collaboration with a domestic airline. Key parameters and possible operational causes were determined based on the expert interviews conducted in the previous study. The Seq2seq method was chosen combined with four types of neural networks: Recurrent Neural Network (RNN), Long Short-term Memory (LSTM), Bi-directional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) to find the best combination for Force and Timing models. The best models for Force and Timing were Seq2seq combined with BiLSTM and GRU respectively, and the best accuracy rates reached 79% and 82% correspondingly. The results verified the feasibility of using machine learning methods to establish an automatic classification system for the operational causes of FOQA events. Methods such as Attention Mechanisms could be taken into account in the future to advance the accuracy of the causal models. The study could contribute to improve the efficiency of the causal analysis for flight abnormal events and to enhance the safety management capability of the air transportation industry.

並列關鍵字

FOQA ； Hard Landing ； Long Landing ； Seq2seq ； Neural Network

參考文獻

外文部分：

1. Ancel, E., Shih, A. (2012). The analysis of the contribution of human factors to the in-flight loss of control accidents. Paper presented at the 12th AIAA Aviation Technology, Integration, and Operations (ATIO) Conference and 14th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 5548.

2. ATM Avionics Division. (2018). QAR Quick Access Recoder.

3. Bao, W., Yue, J., Rao, Y. (2017). A deep learning framework for financial time series using stacked autoencoders and long-short term memory. PloS one, 12(7), e0180944.

4. Boritz, J. E., Kennedy, D. B. (1995). Effectiveness of neural network types for prediction of business failure. Expert Systems with Applications, 9(4), 503-512.