航空安全本是航空公司最重視的環節之一,民用航空運輸業除機組員作業之外,地勤之維修維護作業,亦是重中之重。本研究與國內某民航維修業者合作,運用其歷年危害風險系統紀錄之資料,經由簡易分類建立三種主要事件之危害類別,包括人員、環境以及管理。後續則依據各別事件評估之危害程度運用風險矩陣量化為風險值,再累積建立三種危害類別的每月風險總值,以作為該業者每月風險參考指標。研究中使用四年累積之每月風險指標共50筆,透過類神經網路(Artificial Neural Network, ANN)中的多層感知器(Multi-layer Perceptron, MLP)及循環神經網路(Recurrent Neural Network, RNN)之架構建立三種危害類別之風險預測模型。研究結果顯示,兩種神經網路模型搭配三類危害類別的預測風險值與實際值之相關係數皆介於0.7至0.8之間。雖然仍有樣本數不足及過度描述等問題,但此結果仍驗證了使用類神經網路方法建立風險預測系統的可行性。未來將持續協助業者提升分析效率及預測準確率,節省人力分析及成本,並及時對高危害風險做出預防性處理,促進並落實飛航安全管理,將其安全管理由主動式推向預測式管理系統。
Flight safety is the major operation concern to airlines. Apart from flight operation, aircraft maintenance also plays an important role in the air transportation industry. In this study, we collaborated with a domestic aircraft maintenance organization and used its historical hazard reports to conduct a risk prediction study. The hazard records were classified into three major risk categories: Personnel, Environment and Management based on the context analysis. Every report had been given a risk level during the input process. We used the risk matrix to calculate the corresponding risk value in this study. For each risk category, the risk values belonged to the same month would be accumulated and normalized by the corresponding working hours to establish the monthly risk indicators. The multi-layer perceptron (MLP) model and recurrent neural network (RNN) model were adopted to build the risk prediction models for the three risk categories. The results indicated that the correlation coefficient between the predicted results and the actual risk values ranged between 0.7 to 0.8. Although issues such as insufficient sample size and over-fitting still existed, this study verified the feasibility of using artificial neural network (ANN) methods to establish a risk prediction system in the flight safety field. Practical contributions of this research include providing timely preventive measures to departments with high risks, saving manpower and analysis costs, and enhancing the safety management in a proactive manner.