Mach Number Control Of Continuous Wind Tunnel Based On Machine Learning

Wind tunnel is one of the most commonly used and effective tools for aerodynamic experiments.Maintaining the stability of Mach number in the test section of wind tunnel when changing angle of attack continuously is important for developing high-performance aircraft.At present,there are three commonly used control methods which are PID,model predictive control,and feedforward feedback decoupling control.PID feedback control has a certain time lag and cannot obtain the required control precision of Mach number in the experiments with continuously changing angle of attack.And model predictive control requires solving online constraint optimization problems so that is only applicable to environments with high-performance computing resources.Therefore,feedforward feedback decoupling control is generally used in practical applications.But the original version of this control method has some shortcomings,which are complicated in the experimental steps and low in control accuracy.To overcome these defects,in this thesis,we propose to clean and model wind tunnel experimental data using machine learning algorithm to optimize the control process.The main research work of this thesis is as follows:(1)In view of the stationary single kernel Gaussian process has less long-range structure,this thesis proposes to use additive Gaussian process to fit wind tunnel data based on the idea of multiple kernel learning.As a result,a new control framework is put forward,which takes the trained model as the feedforward part of the Mach number control system.In order to verify the effectiveness of the additive Gaussian process,this thesis applies two different strategies to divide data set to train the model and compares the trained model with traditional regression models as well as single kernel Gaussian process.The experiment results show that additive Gaussian process has the best accuracy.(2)When a data-driven method is employed for Mach number control,the larger the amount of data,the higher the accuracy of model.However,the acquisition cost of wind tunnel data is very high,and it is often difficult to obtain a large amount of data in practical applications.For purpose of using less experimental data to achieve a satisfactory control accuracy,in this thesis,a Bayesian guided method is proposed to acquire new data.The experiment results verify the effectiveness of this method.(3)Due to the influence of operational errors or other unknown factors,some abnormal data generated during the wind tunnel experiment,which will have a negative effect on model fitting.Therefore,these abnormal data need to be detected and cleaned.Cleaning data manually is time-consuming and is not practical when the data set is large.Hence,based on the characteristic of wind tunnel data acquisition in chronological order,this thesis applies Prophet,which is a model for time series,to conducting anomaly detection.In addition,in order to alleviate the impact of severely abnormal data on Prophet fitting,an anomaly detection framework that combines coarse and fine is proposed.Experiments show that this anomaly detection framework has high recall and AUC.