Research On Aeroengine Thrust Estimation Methods Based On Neural Networks

Aeroengine provides the thrust for the aircraft,and it's one of the most important components of the aircraft.Performance parameters,such as thrust and stall margins,provide crucial information for operating an aeroengine in a safe and efficient manner.As a result,thrust regulation is often one of the primary goals in engine control.Since the thrust of the aeroengine is unmeasurable in the flight,the traditional control method usually takes the parameters tightly related to the thrust such as the rotor speed and pressure ratio as the input signals to control the thrust indirectly.However,to protect the aero-engine,this method reserves enough safety margin at the price of the performance compromise of the aeroengine.A better way which can fully utilize the potential of the aeroengine is to control the thrust directly if the thrust can be estimated.Based on the consideration above and the existing studies,the author proposed several algorithms to estimate the aeroengine thrust by combining artificial intelligent optimization with artificial neural network(ANN),which obtained satisfactory results.The main contents of this thesis include as follows:Firstly,to obtain a better thrust estimator based on ANN,the author modified and improved particle swarm optimization(PSO)to adjust the network size of ANN dynamically and optimize ANN parameters simultaneously.In the proposed algorithm,the author randomly initialized the network's spreads and calculated the connection weights of ANN by Moore-Penrose inverse instead of optimizing all these parameters using PSO.In this way,not only the dimension of the parameters required to be optimized was reduced,but also the performance of ANN could be well adjusted by the Moore-Penrose inverse.This algorithm in this thesis was referred to as MGPSO-SORBF.Secondly,the shortcomings of MGPSO-SORBF were analyzed at first such as the network size was hard to jump out from the optimal network size after the predetermined iterations number.After the analysis,a new algorithm combined the softmax function to adjust the network size was proposed.In the new algorithm,one or more network sizes would be removed from the predetermined set after each predetermined iteration epoch.The particles related to the removed network sizes would be allocated to the other small groups whose network sizes were not removed.Besides,the new algorithm developed a new method based on PSO algorithm to solve the high-dimensional optimization problem.This method improved the estimated accuracy of the estimator based on ANN during the whole optimization process.In this thesis,the new algorithm was named as the HDPSO-STRBF algorithm.At last,the estimation method based on the LSTM algorithm was proposed to meet the requirements of the thrust estimation of the aeroengine on the transition state.To improve the estimated accuracy,the author took the LSTM network as the base estimator and combined it with the gradient boosting,which came into the proposed LSTM-GB algorithm.The estimator based on the LSTM-GB algorithm can estimate the thrust of the current time by several aeroengine states before the current time.The experimental results showed that the estimator based on the LSTM-GB algorithm could achieve satisfactory results.