Using Machine Learning to train a Digital Test Pilot for missions in turbulent environments

虽然在全运动模拟器中的有人驾驶飞行模拟可以用于通过支持海上试验来降低成本，但是模拟试验仍然昂贵并且受到模拟器可用性的限制。该项目的目的是开发一个人类飞行员的数学模型，一个数字试飞员，它可以用来进行多个虚拟甲板着陆，以建立安全操作包线的可能边界。飞行员将使用测试飞机在实际海上试验中收集的大量未开发数据进行“训练”。利用这些数据的能力可以获得巨大的成本效益，并在优化设计，维护和使用方面进一步提取能力。该项目将研究“智能”使用和融合可用数据的方法的适用性，以开发数字试飞员。基于桌面的预测模拟工具，在综合的驾驶员-车辆-环境中使用客观优化的人类驾驶员建模技术。多回路追踪飞行员模型，使用线性化的直升机飞行动力学模型与空间空气湍流模型，已显示出潜在的使用作为一个预测工具的操作间隙。在最近与英国国防部和Nova系统公司的讨论中，提出了一种更先进的建模方法，即采用一种非线性直升机飞行模型，该模型具有一个飞机可以随意“探索”的无限制紊流空气流场。其目的是使用机器学习和来自飞机状态和飞行员控制输入的数据来“训练”数字试飞员模型，这些数据来自海上试验和在大学自己的HELIFLIGHT-R模拟器内进行的模拟飞行试验。数字试飞员模型训练将涉及融合来自一系列输入源的数据，例如飞机状态、飞行员控制输入、船上运动，以确定适当的响应。Nova准备为该项目提供资金支持，而国防部也表示他们准备提供真实世界的数据;两者都热衷于提供专业知识。人们还认识到，该过程可以应用于非军用船舶/直升机操作，以及在海上平台和医疗和救援服务中操作的直升机。该项目还将审查这些进一步利用的途径。

Whilst piloted flight simulations in a full-motion simulator can be used to reduce costs by supporting the at-sea trials, the simulated trials are still expensive and are limited by simulator availability. The aim of this project is to develop a mathematical model of a human pilot, a Digital Test Pilot, which can be used to conduct multiple virtual deck landings in order to establish the likely boundaries of the safe operational envelope. The pilot will be 'trained' using the wealth of untapped data that is gathered by the test aircraft during the real-world sea trials. The ability to exploit this data could reap large cost-benefits and extract further capability in optimised design, maintenance and usage. The project will examine the suitability of methods for 'intelligent' use and fusion of available data to develop the Digital Test Pilot.Desktop-based predictive simulation tools that use an objectively optimized human pilot modelling technique within an integrated pilot-vehicle-environment are available. A multi-loop pursuit pilot model, using a linearized helicopter flight dynamics model with a spatial air turbulence model, has shown potential for use as a predictive tool for operational clearances. In recent discussions with the UK MoD and Nova Systems, a more advanced modelling method has been proposed to use a non-linear helicopter flight model with an unrestricted turbulent air flow-field which the aircraft can 'explore' at will. The intention is to 'train' the Digital Test Pilot model using machine learning and data from the aircraft states and pilot control inputs derived from at-sea trails and from simulated flight trials undertaken within the University's own HELIFLIGHT-R simulator. The Digital Test Pilot model training will involve fusing data from a range of input sources e.g. aircraft state's, pilot control inputs, shipboard motion, to determine an appropriate response. Nova are prepared to financially support the project, and MoD have indicated they are prepared to provide real-world data; both are keen to provide expertise.It has also been recognised that the process could be applied to non-military ship/helicopter operations as well as helicopters operating to offshore platforms and those in medical and rescue services. These avenues of further exploitation will also be examined in the project.