The Research Of Ensemble Methods Based On Big Data For The Mach Number Prediction In Wind Tunnel

A wind tunnel is a device of aerodynamic measurements for the beginning stage of the aircraft design.The Mach number in the test section is an important performance index for measurements,and the stability of it has a significant effect on the quality of the flow.To realize the precise control,the Mach number prediction is requied to be fast and precise.However,the big data accumulated from measurements,with the characteristics of large-scale samples and high-dimensional input features,et al.,is the main issue to predict the Mach number fast and precisely.This paper focuses on the research of ensemble methods based on the big data for the Mach number prediction in a wind tunnel.The main works are as follows:1.The structure of the flow model in a wind tunnel is constructed.Firstly,the FL-26 transonic wind tunnel is selected as the research object.To provide experience knowledge for building a data-driven model,the properties of the pneumatic structure,the flow and the working conditions of measurements are analysed.Secondly,the structure of the Mach number data-driven model is constructed,and the NARX representations for the Total Pressure(TP)and the Static Pressure(SP)are gained respectively.Then,the main issue to realize the data-driven Mach number model is analysed,and the scheme of solving the issue is to build ensemble models.2.Ensemble methods for a Mach number prediction model are studied.For ensemble models,constructing independent sample subsets or feature subsets are effective ways to reduce the complexity and to solve the issue of building models on big data.Firstly,the random forest Mach number model is built.To reduce the complexity,the random forest is the most popular sample subsets based ensemble method.Experiments show that,with two working conditions,the random forest Mach number model obtains satisfying performance,but with three working conditions,the forecasting speed and precision of it reduce obviously.The main reason is that the random forest cannot solve the high-dimension issue.Additionally,the scale of samples becomes larger,the nonlinear stronger&the distribution of the data unbalanced.As the measurement data get more complexity,the random forest cannot meet the requirements of the forecasting speed and the precision.Then,based on the multivariate fuzzy Taylor theorem,the Feature Subsets Ensemble(FSE)method is proposed.In an FSE model,the feature-space is divided into low-dimensional feature subsets directly,quickly and exhaustively.The reduction of the dimensionality helps feature subsets to save more memory space in computers.It indicates that "smaller-scale" and "more balanced" training sets are used,which makes the Mach number prediction fast and precise.The FSE method can solve the large volume,the variety and the low velocity issues of the big data effectively.3.For big data sets,base learners,i.e.learning algorithms of sub-models,in an FSE model are studied.Precise and diverse sub-models are necessary to design good ensemble models.It is important to have a suitable base learner which helps the FSE method solve the low value issue of the big data.Thus,the effectivenesses of the FSE for unstable learners and stable learners are studied,and the BP network and the fixed-size LS-SVM are the examplers,respectively.Experiments show that,for the Mach number prediction under three working conditions,compared with a single model,the FSE model can reduce the complexity dramatically,improve the precision,and meet the requirements of the forecasting speed and the precision.With the FSE structure,for the forecasting speed and precision of the Mach number prediction:as unstable learners,non-linear strong learner BP network outperforms linear weak learner regression tree;and as nonlinear strong learners,the fixed-size LS-SVM outperforms the BP network.4.The ensemble pruning method for the FSE model is studied.Too many and redundant sub-models may limit the application potential of FSE models.Thus,it is necessary to reduce the number of sub-models,i.e.to make the ensemble pruning,with the same precision of the entire ensemble.For this issue,the Maximum Entropy based Pruning(MEP)method is proposed.Firstly,all sub-models are ordered with errors,from the smallest to the largest.The P*sub-models with the lowest errors are selected as an original working set.Secondly,the maximum entropy of a working set is used as the criterion to replace a sub-model from the former working set.In the MEP method,both the precision and the diversity of sub-models are considered.The MEP-FSE method can solve the variety and the low value issues of the big data for a certain extent.Experiments show that the MEP-FSE Mach number prediction model outperforms the FSE model on the precision.5.The robustness of the FSE method on the noise data is studied.In an FSE model,the feature-space is splited exhaustively and the entire training set is used at least once,which may limit the precision improvement of an FSE model.To improve the robustness of the FSE model,on feature subsets the Bootstrap sampling approach is introduced to limit the reuse of noise data,and the Bootstrap-FSE method is proposed.The Bootstrap-FSE method can solve the low value issue of the big data effectively,and solve the low velocity issue for a certain extent.Then,the Bootstrap-FSE Mach number model is built,and is tested on both the low and the high noise data sets.Experiments show that the robustness of the Bootstrap-FSE model is better than that of the FSE model.The Bootstrap-FSE model can further improve the forecasting speed and precision of the Mach number prediction.