Design Of Control Laws And Intelligent Optimization Of Variable Cycle Aero Engine

Compared to conventional aero engine,variable cycle aero engine has the advantages of more variable geometry variable structure,more complex structure of the system itself,more adjustable components,low fuel consumption and high thrust,etc.But it also has the characteristics of wide range of flight envelope,strong non-linearity,time variability and uncertainty.This paper takes a variable cycle aero engine as the object of research.Firstly,the engine control law is designed and a deep reinforcement learning intelligent algorithm is introduced to solve for the lowest fuel consumption at steady state and the maximum thrust at transition state as well as the maximum thrust output in the shortest time under the transition state of the variable-cycle engine.Finally,the incremental learning method is used to achieve autonomous engine development in multiple modes with less training data,which improves the control adaptability of the engine in different modes,and the main research contents are as follows:Firstly,a model of the variable cycle engine mechanism is developed to simulate the operation of the variable cycle engine,including three stages: start-up,cruise and end.It also analyses the problem of finding the best performance of variable cycle engine.Secondly,to address the problem of designing the steady state control law for variable cycle engine,according to the characteristics of variable cycle engine with many control parameters and state parameters,seven variable geometry variable structures of variable cycle engine are selected as the control inputs and the high-pressure rotor speed and fan pressure ratio as the control outputs according to the conventional variable cycle engine parameter selection convention to design a 7-input,2-output control law.While using deep reinforcement learning algorithms to construct an algorithmic State space,Action space,and Reward function,with the objective of pursuing the lowest fuel consumption at steady state,two deep reinforcement learning algorithms,DQN and DDQN,are used to optimizing the control law.The simulation results show that,under the premise of ensuring that the engine does not over-temperature,wheeze or over-rotate,the deep reinforcement learning algorithm can achieve arbitrary non-linear mapping by relying on its powerful fitting ability,and is not affected by the model accuracy and the coupling relationship between the non-linear variables,which is well achieved to achieve the lowest fuel consumption at a steady-state point of the variable cycle engine.Thirdly,for the problem of designing the transition state control law for variable cycle engine,the same 7-input,2-output control law is designed and the control law is optimized using two deep reinforcement learning algorithms,DQN and DDPG.Simulation comparison results show that the above method relies on multilayer neural network iterative optimization,achieving the goal of optimizing the maximum thrust output of a variable cycle engine in the shortest possible time in the transition state,and the engine does not over-temperature,wheeze or over-rotate.Finally,a variable cycle engine control law model is developed using a radial basis function(RBF)model.The ALD-L2 KRLS algorithm is used as an incremental learning algorithm to give the variable cycle engine control law model the ability to incrementally adapt to different modes.A fast and accurate mapping model between engine control parameters and state parameters is obtained,and the model is allowed to adapt to new modes without forgetting the old ones.