Relief和SVMRFE在高超声速进气道不起动预测中的应用

Application of Relief and SVMRFE in Predicting Unstart of Hypersonic Inlet

高超声速进气道不起动预测研究中主要包括确定压力传感器位置和建立起动/不起动分类面,属于机器学习中特征选择问题和分类问题,而常用特征选择算法(基于支持向量机的递归特征消除SVM-RFE)单一并且耗时较长;为解决该问题寻找较优的特征选择算法,建立一个高超声速二元进气道/隔离段模型,通过数值模拟获得内流道上表面压力数据样本;利用Relief和SVMRFE组合式算法Relief-Corre方法,Relief-SVMRFE方法,Relief-PSO-SVMRFE方法进行特征选择;支持向量机SVM训练分类面;最后得出Relief-SVMRFE方法性能最优,运行效率比SVMRFE提高了约3倍,准确率比其他基于Relief组合方法高;获得最优特征的分类面具有较高的泛化性与鲁棒性,证明该分类面的有效性。

The predicting unstart of hypersonic inlet consists of calculating positions of pressure sensors and establishing classification of inlet start/unstart,the feature selection algorithm and the classification algorithm was taken to solve this two problems,the common algorithm(SVMRFE)is time intensive on computation.A 2D hypersonic inlet/isolator model were simulated to generate wall static pressures data set.Hybrid feature selection algorithm based on SVMRFE algorithm and Relief algorithm were used to select optimal pressure points,which were called as Relief-Corre algorithm,Relief-SVMRFE algorithm,Relief-PSO-SVMRFE algorithm.The support vector machine(SVM)algorithm was used to train the classification plane.Finally,the performance of Relief-SVMRFE algorithm is proved best,since its operation efficiency is three times higher than SVMRFE and it has higher accuracy than other algorithms based on Relief.The hyperplane with strong robust performance and generalization performance conforms to the actual physical law,so the result shows that the criterion is valid.