Automated UAV Mission Optimisation Using an On-Board Digital Twin for Long Range Wildlife Conservation Tasks

点击翻译按钮获取中文摘要

The University of Bristol Flight Lab is extending its focus to long range UAV operations for animal conservation as part of the WildDrone Marie Curie Network grant starting in January 2023. This brings together expertise in beyond visual line of sight (BVLOS) operations, aircraft control and aircraft design to achieve extremely demanding missions. In parallel, the Flight Coach Project is working to automate analysis of precision aerobatic flights with advanced flight data post-processing capabilities. This work brings the two projects together to optimise fixed wing Uninhabited Aerial Vehicle (UAV) missions based on continuous, automated analysis of flight data. Significant work exists in the development of highly efficient UAVs as well as in tools to predict their performance. These UAVs operate at relatively low flight speeds in relation to the local wind speeds, meaning that their potential mission performance is strongly related to the chosen operating parameters. Despite this, operational decisions for UAVs often take only limited consideration of the performance model and the local environment. This research aims to close this gap, automating the performance model estimation and wind estimation to feed into real time mission parameter and flight path optimisation. Part of the process developed as part of the Flight Coach project includes collecting and labelling flight data using a temporal alignment algorithm. This process is invaluable for Flight Regime Recognition (FRR). A variety of methods are used for FRR, one of which is to train a neural network with labelled training data to predict the regime based on state observations. Other research shows that FRR using low cost flight data recorders is an effective method of estimating loads on aircraft components, and can even be preferable to directly mounting strain sensors on the component being analysed. The trained FRR agents developed based on this process will be deployed on-board to automatically recognise steady state and manoeuvring flight regimes. Real Time Parameter Identification (RTPID) and model estimation are well researched areas and various techniques already exist. These methods will be reviewed, selected and deployed on a representative test bed, with a view to standardising the processes and systems integration as an add on to the Ardupilot open source autopilot system. The potential utility of this process deployed on an onboard companion computer extends beyond the mission and performance optimisations described above, especially to enhancing safety of UAV operations by detecting failures earlier. To produce reliable mission optimisations, estimates of the environmental conditions in the projected airspace must be produced. Local, on-board wind estimation can be achieved for a UAV by utilising the continuously updated aircraft parameter model and IMU measurements. Methods of extrapolating low level measurements to larger height ranges will also be developed. The novel contribution of this work in this area is to implement and couple these two tasks to provide an improving projected wind field as the mission progresses. The aim of this project is to develop automated methods to optimise fixed wing UAV mission performance based on flight data observations. The following objectives are proposed:-Identify key flight regimes using neural networks trained on automatically labelled flight data during steady state flight and rapid manoeuvres.-Define and automate a real time parameter identification (RTPID) process, coupled with live wind field estimation to provide a continuously updated on-board digital twin.-Apply the digital twin and environment estimate to real time mission parameter optimisation to minimise energy usage in long range surveying missions.-Extend the scope to consider specific individual animal identification tasks, to automate path generation and course...