光谱分析中的支持向量机方法及其性能优化

Support Vector Machine and Optimized Method for Spectral Analysis

针对红外光谱气体分析中建立数据模型需要标定大量样本的问题,提出一种基于正则理论的支持向量机的小样本机器学习方法,该方法能在获得模型参数全局最优点的同时保证训练误差为零,因而能较好地消除光谱间的交叉敏感现象,利用其良好的非线性映射能力对多组分红外光谱仪的试验结果表明,该方法可使光谱仪的交叉灵敏度下降约81倍。针对支持向量机(SVM)没有足够的理论支持的结构参数选取比较困难的问题,提出一种基于遗传算法和交叉检验相结合的遗传支持向量机(GA_SVM)算法,利用遗传算法的随机搜索特性求取SVM的最优结构参数,在20世代即可求取光谱仪的最小均方根误差(MSE)0.018,并且在算法的前数世代,系统的MSE即已开始成倍下降。这些结果表明GA_SVM光谱仪具有更高的效率和泛华能力。

According to support vector machine based on the regularization theory, a small scale machine study theory was proposed to solve the problem of multi-gas analysis, which is mainly restricted by the lack of experimental samples. With its well nonlinear mapping ability, the training error was decided to be zero and global optimal parameters were obtained, hence the cross-sensitivity of spectrum is preferably eliminated. In multi-component gas analysis, the results show that the cross-sensitivity decreased to 1/81. A method based on genetic algorithm and cross-validation was proposed to solve the parameters selection of support vector maehlne（SVM）, which still lacks theory support. The optimal structure parameters were achieved by genetic random search algorithm, the mean square error（MSE）0. 018 of the spectrometer was achieved in 20th generation, and MSE decreased by multi-times in the fore generations. This hints that the genetic algorithm SVM is more efficient and has better generalizing ability.