基于LS-SVM的苹果近红外光谱回归模型的研究

Regression Model for Apples' Near Infrared Spectroscopy Based on LS-SVM

提出了一种应用于苹果近红外光谱分析的LS-SVM回归模型;使用基于两层网格搜索的交叉验证算法确定LS-SVM的最优参数(γ,σ2),缩短搜索时间并提高了效率;针对LS-SVM稀疏性的缺乏和鲁棒性的不足,对模型进行优化训练。在不同方差的噪声下,通过优化训练模型的抗干扰能力明显强于常规训练模型;将优化后的LS-SVM模型应用于苹果酸度的预测,利用光纤光谱仪采集苹果近红外吸收光谱作为模型输入,使用酸度计测得苹果测量酸度值作为模型输出;实验结果表明,所建模型的相关系数和均方根误差为0.9615和0.0312,与MLR、PLR、ANN和常规LS-SVM模型比较,优化后的LS-SVM具有更好的回归性能。

A regression model based on Least--squares support vector machine （LS-SVM） was developed tor analyzing the apples＇ near infrared spectroscopy. Cross- validation algorithm based on two--grid search, which was put forward to determine the optimal parameters （γ, σ2） of LS--SVM, shortened the search time and enhanced efficiency. Due to the lack of sparseness and robustness, the optimization training was carried out. Under the noise with different variances, the stability of the model by optimization training was superior to that by standard training. This model was applied to forecast the acidity of the apples. The near--infrared absorption spectra of the apples were collected by the Fiber Optic Spectrometer as the input of the model, meanwhile the actual acidity of the apples were measured by the PH meter as the output. The experimental results showed that the correlation coefficient R2 was 0. 9615 and root--mean--square error RMSEP was 0. 0312 obtained by the above model. Moreover, the comparison of the model of the MLR. PLR, ANN and the conventional LS--SVM indicated that the LS--SVM with optimization had the higher level of regression performance