基于优化支持向量机的玉米淀粉含量估计

Corn Starch Content Estimation Based on Optimized Support Vector Machine

为了优化回归模型参数和提高模型的预测性能,建立了一个基于粒子群算法优化支持向量机的回归预测模型,实现了对玉米淀粉含量的有效估计。研究基于公开的玉米光谱数据集,首先分别采用源光谱、标准正态变量变换(Standard Normal Variate transform,SNV)、SNV+SG (Savitzky-Golay)卷积平滑算法、MSC(Mutiplicative Scatter Correction,MSC)多元散射校正、MSC+SNV、一阶求导(First order derivation,FD)、MSC+SNV+FD方法对光谱数据进行预处理,去除数据噪音等冗余信息;其次使用主成分分析(Principal Component Analysis,PCA)算法进行高维光谱数据的特征提取,获得数据的有效特征;最后结合粒子群优化算法(Particle Swarm Optimization,PSO)参数寻优支持向量回归(Support Vector Regression,SVR)建模中的重要参数,实现对玉米淀粉含量的预测建模。实验对比不同预处理方法、PSO优化过程中不同粒子数时的模型预测效果。实验结果表明:MSC+SNV预处理时,PSO算法中粒子数参数为60时,训练集的RMSE为0.1 754,R2为0.9 583,预测集的RMSE为0.2 036,R~2为0.8 863,预测相对分析误差RPD为3.0 631,模型具有很好的预测效果。

For optimizing the parameters and improving the prediction performance, a regression prediction model using particle swarm optimization algorithm to optimize support vector machine is established, which achieves the effective estimation of corn starch content. In this paper, based on using the open corn spectral data set, firstly, the spectral data are preprocessed by using source spectrum, standard normal variate transform(SNV), SNV+SG(savitzky-Golay) convolution smoothing algorithm, multiplicative scatter correction(MSC), MSC +SNV, first order derivation(FD), MSC +SNV +FD method separately to remove redundant information such as data noise;secondly, principal component analysis(PCA) algorithm is used to extract the features of high-dimensional spectral data to obtain the effective features of the data;finally, combining with particle swarm optimization(PSO) to optimize the important parameters of support vector regression(SVR)modeling is included, so as to accomplish the prediction modeling of corn starch content. The experiment compares the prediction effect of the model with different pretreatment methods and different particle numbers in the PSO optimization process. And the results show that when the conditions are MSC+SNV preprocessing and 60 particle numbers, the RMSE of training set is 0.1 754, R2is 0.9 583, the RMSE of prediction set is0.2 036, R2is 0.8 863, and the prediction RPD is 3.0 631,which means that the model has a good prediction effect.