Aerodynamicmodeling using an end-to-end learning attitude dynamics network for flight control

A novel identification method of aerodynamicmodels using a physics neural network,named the attitude dynamics network,which incorporates the attitude dynamics of an aircraft without any prior aerodynamic knowledge,is proposed.Then a learning controller,which combines feedback linearization with sliding mode control,is developed by introducing the learned aerodynamicmodels.The merit of the identification method is that the aerodynamicmodels can be learned end-to-end by the physics network directly from the flight data.Consequently,the paper uses an offline scheme and an online scheme to combine the identification process and the control process.In the offline scheme,learning the aerodynamic models and controlling the aircraft compose a cascade system,whereas the online scheme,similar to Learn-to-Fly,is a parallel system.Specifically,in the offline scheme,the physics neural network is trained by sufficient offline flight data,and then the trained network is substituted into the controller.The online scheme refers to the controller making the aircraft fly to generate flight data and sending these data to the deep network at the time of training,while the deep network provides the trained aerodynamic models to the controller at other times.Simulation results show that both under nominal and disturbance aerodynamic conditions,the network trained offline with a large amount of nominal data approximate the aerodynamicmodels well.Thus,the performance of the controller reaches a good level;for the online scheme,the predictive capability of the network increases and the performance of the controller improves with more training data.