多旋翼無人機地面風險之評估

Abstract

本研究透過無人機墜機可能性推論模型評估地面風險。相對於有人機，由於無人機的取得容易和操作門檻較低，使其快速普及並廣泛應用；加上無人機可在低空飛行且數量急遽增加，這使得地面風險大幅提高。過去的研究著重於某些因素對地面風險的影響，本研究根據前人研究成果進行全面且系統化地評估。地面風險綜合考量墜機發生率、機體衝擊與動態人口密度。其中墜機發生原因包括人為操作、無人機本身的抗風能力和風況、飛航中的能量損失、衛星導航問題以及電磁干擾引起的通訊失效。利用文獻研究、數據、公式以及機器學習工具分別建立人為操作、抗風、能量損失、衛星通訊以及電磁干擾五項墜機推論子模型。研究中比較隨機森林和深度神經網路之差異，並選擇適用的機器學習方法。綜合考慮動態人口密度和機體衝擊所致之傷亡率，計算每平方公里無人機造成的地面風險值。結果顯示使用深度神經網路可以建立一個不斷更新且更接近實際情況的預測模型。未來無人機飛航管理系統若成功地建置，即可收集大量的真實數據，建立一個數據驅動之地面風險評估系統，如此可以提供監管決策、飛航路徑規劃、地面風險地圖建置之參考依據。

This study established a model to predict the probability of unmanned aerial vehicle (UAV) crash and conduct ground risk assessment. The greater availability and simpler operation of UAVs compared with manned aerial vehicles have contributed to their growing prevalence and extensive applications among the public. However, the ability of UAVs to fly at low altitudes and the rapidly growing number of UAVs both contribute greatly to the increase in ground risk. Most studies have discussed the effect of few factors on ground risk. To address this research gap, this study conducted a comprehensive and systematic review of relevant literature and accounted for crash probability, UAV impact, and population density in ground risk assessment. Reasons of UAV crashes may concern human operation, wind resistance of UAVs, wind conditions, energy loss during flight, satellite navigation problems, and lost communication due to electromagnetic interference. Accordingly, relevant literature findings, data, equations, and machine learning techniques were employed to establish five crash probability submodels, which were individually based on human operation, wind resistance, energy loss, satellite navigation, or electromagnetic interference. Prediction results obtained from the random forest regression algorithm and deep neural network algorithm were compared to identify the optimal machine learning model. Ground risk attributed to UAV operation per km2 was calculated based on population density and mortality and injuries due to UAV impact. The results verified that the deep neural network model yielded more accurate prediction results through constant update, thereby more closely reflecting real-world situations compared with the random forest model. In sum, the construction of a UAV flight management system enables the collection of immense real-world data, which can be used to create a data-driven ground risk assessment system and thereby provide referential data for regulatory decision-making, flight route planning, and ground risk mapping.