應用循環神經網路建立航空器落地常發事件之肇因模型

Establishing a Causal Analysis Model of Aircraft Frequent Landing Events by using Recurrent Neural Networks

黃哲暐(Che-Wei Huang)

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(2020年)

https://doi.org/10.6840/cycu202000759

摘要

航空安全一直以來都是航空運輸業者的最注重的核心。對於飛行途中發生的不正常狀況，民航業者多使用快速擷取記錄器(QAR)記錄與分析不正常航班參數超限情形，例如降落力道過重或飛機平飄過久等，以做為後續機隊飛航訓練、運作改善與維修保養之依據。但是目前需要仰賴專家人工判讀，找出事件發生時參數超限背後的肇因，包括飛行員操作不足或外界環境等影響因素。本研究目的在於運用兩種類神經網路法建立特定常發事件之判讀模型，藉由使用人工智慧方法，分析QAR中的重要參數，推測出事件可能操作肇因，來提高事件判讀效率。分析數據來自國內某民航業者所提供之兩種落地常發事件之參數資料，再經由專家訪談找出分析時使用的重要參數與可能肇因之後，使用長短期記憶(Long Short-Term Memory, LSTM)和雙向長短期記憶(Bi-directional Long Short-Term Memory, BiLSTM)以建立事件操作肇因分類模型。研究結果顯示運用雙向長短期記憶所建立之分類模型在正確率上可達到七成，雖然因實務研究無法避免樣本數過少、資料不平衡以及過度擬合等問題，但此結果仍驗證了使用類神經網路方法建立不正常事件操作肇因之自動判讀機制的可行性，未來可繼續在此領域深入研究探討，以促進飛安預防式管理之精進。

關鍵字

QAR ； 不正常事件 ； 類神經網路 ； 長短期記憶 ； 雙向長短期記憶

並列摘要

Flight safety is the foundation of the air transportation industry. For anomaly situation during flight, the operators would use the Quick Access Recorder (QAR) to detect parameters exceedance and to conduct the following causal analysis. The possible corresponding improvements include actions such as follow-up training, procedure modification or preventive maintenance. But most companies still rely on subject matter experts (SMEs) to do the analysis to determine the causes behind the anomalies, including pilot operational causes and environmental factors. The purpose of this study is to establish a cause classification model for frequent anomalies using neural network methods to improve the efficiency of event interpretation. The input data are the QAR parameters, and the outputs are the possible causes. The dataset comes from the QAR data of two specific frequent occurrences provided by a domestic company. Based on the discussion of expert interviews, the key QAR parameters and possible causes were defined. Long Short-Term Memory (LSTM) and Bi-directional Long Short-Term Memory (BiLSTM) were adopted to establish the classification model of operation cause for abnormal events. The result proved that both LSTM and BiLSTM could achieve an accuracy higher than 70%. Although issues such as sparse sample size, imbalance and over-fitting still existed, the result validates the practices of neural network methods to establish AI causal classification of flight events.

並列關鍵字

QAR ； Neural Network ； LSTM ； Bi- LSTM

參考文獻

1. Smithsonian Institution. (2010). The Wright Brothers The Invention of the Aerial Age

2. International Air Transport Association. (2019). Safety Report 2018 Issued April 2019 Edition55.

3. FAA. (2004). AC 120-82 - Flight Operational Quality Assurance.

4. McCulloch, Warren S.; Pitts, Walter. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biology, 5(4), 115-133.

5. Douglas Eck.; Jurgen Schmidhuber. (2002).Finding temporal structure in music: blues improvisation with LSTM recurrent networks