基於卷積神經網路發展無人機飛航風險評估系統

Abstract

本研究針對無人機飛航風險發展一套數據驅動之評估系統。隨著無人機應用越來越廣泛且價格逐漸親民，使用頻率快速增加，惟因無人機較有人機的飛行空域低且重量輕，使得飛航安全、安寧與隱私等問題逐漸受到普遍的關注，在此背景下勢必需要一套考慮環境與天候對無人機飛航安全影響之系統。本系統係屬團隊研究的一部分，整體研究除了針對空域中之無人機進行風險評估外，同時亦對無人機進行分散式監管，並以收集到的飛航數據利用區塊鏈架構進行個人化保險費率之計算。由於無人機較易受到天候、地形及訊號等因素的影響，可能導致無人機飛行偏差甚至發生事故，因此飛航風險評估在無人機監管中將扮演不可或缺的角色。本研究開發一套無人機飛航風險評估方法，利用大量歷史飛行數據配合數值地表模型建立風險模型以預測不同情境下之風險分佈。惟因現階段實際飛行數據不足，本研究根據有限的實際飛行數據，透過WindPerfect風場模擬軟體產生大量的模擬數據，並以卷積神經網路歸納分析風場、衛星定位及通信訊號強度與飛行路徑偏差間的交互關係，進而建立無人機飛航風險等級圖，風險等級地圖包含城市、山區與海岸等場域，大致包含目前無人機操作者作業之情境，利用色階深淺表示風險的高低程度，並設計親善的使用者介面，提供操作者進行風險評估與任務規劃。透過無人機飛航風險等級圖，操作者可於起飛前初步得知各區域之風險，以針對飛航任務進行相應的調整，達到降低飛航風險之目的。

This study develops a data-driven system for the risk evaluation of unmanned aerial vehicles (UAV ) flights. UAVs’ prices are getting more affordable and growing in popularity. However, due to flying at low attitudes and light weight, the security and privacy issues it derives are of widespread concern. Therefore, it is urgent to develop a UAV flight risk assessment system which is part of an overall study by our UAV research team. In addition to risk assessment, the entire research topics include decentralized supervision of UAVs and personal insurance premiums calculation based on historical flight data and claims records. UAV flights are easily affected by wind, terrain, navigation and communication quality etc. which led to deviate from planned flight courses or even crash. This study uses lots of historical flight data and digital surface model to establish a risk model for constructing risk maps in different scenarios, including various wind speed, communication quality, urban, mountainous, and coastal areas. Due to the lack of flight data at this stage, our study generates wind field simulation data using the WindPerfect software, GPS/communication signal quality and flight deviation based on actual flight experience. With those data, a convolution neural network (CNN ) model is established to construct the flight risk map. A friendly user interface is also developed to provide operators to plan flight paths and avoid high risk area based on the UAV flight risk map before takeoff.