LightGBM Algorithm For Compressor Flow Control Logic In Civil Turbofan Engines

The airflow control system of aero-engine compressor is one of the most important parts of the engine.It ensures flight safety by avoiding dangerous conditions such as rotating stall and surge of the compressor,and improves the performance of the compressor.The two important subsystems,Variable Stator Vane(VSV)and Variable Bleed Valve(VBV),are the major section of the compressor flow control system.By adjusting the angle of VSV and the opening of VBV,they stabilize the gas flow status in the compressor air passage,providing necessary conditions for the stable operation of the engine.In civil aviation,accidents threatening aircraft flight safety caused by compressor airflow control system failure often occur.However,as a non-OEM(Original Equipment Manufacturer)airline,it is difficult to study and understand the logic and design concept of the compressor flow control system deeply,and to avoid accidents caused by compressor flow control failure in actual flight.There is also less discussion and analysis on prediction direction in previous compressor related research,but predicting the change trend of compressor flow control system has a great contribution to the early warning of faults.Therefore,it is very necessary to propose a new way to study the compressor flow control logic.In order to study the compressor flow control logic of civil turbofan engine deeply,this thesis uses the Light GBM machine learning algorithm,which is advanced in the industry,to study the baseline model and trend prediction of the compressor flow control logic based on the real operational parameter data of engines collected by the quick access recorder(QAR).Unlike most previous studies that use simulation data or open source datasets,this thesis collects million-level data of different flights by QAR and processes them to form high-quality datasets,in order to improve the model performance and close to the actual engine operation state.Two different types of civil turbofan engines are studied to verify the applicability of the method.The optimization results of the greedy grid search algorithm are used as the hyperparameter settings of the VSV,VBV baseline and prediction model to further improve the model performance,and evaluated by the root mean square error(RMSE)and mean absolute error(MAE).To facilitate the error comparison,we perform a normalization-like process on MAE.The experimental results demonstrate that,under the condition of reasonable input features and sufficient data,high-accuracy baseline and prediction models for VSV and VBV control systems can be achieved,with MAE error ratio not exceeding 0.01.In addition,we propose a sliding window method for short-term prediction error analysis and prediction accuracy reference,that is,accumulating the errors of five predictions for early warning of error accumulation in practical applications.When the condition is satisfied,the MAE sliding window error ratio of the prediction model proposed in this thesis is about 0.02,showing a high prediction accuracy.All of these are based on randomly extracted input datasets from both transient state and steady state data to achieve the purpose of giving baseline and prediction models some generality.To verify the model generality,we finally input transient state and steady state data separately to validate the accuracy of baseline and prediction models under both states.We also input the whole test dataset according to the time sequence of aircraft flight to simulate the output performance of the model in actual operation.This avoids the problems such as inconsistent division standards,difficult state identification and automatic switching of sub-models that are not easy to implement in actual industrial applications caused by dividing research models according to flight segments and states and establishing subsystems in most previous related studies.