遗传算法在鱼粉中肉骨粉近红外光谱检测中的应用

Genetic Algorithm Used for Predicting Meat and Bone Meal Content in Fishmeal by Near Infrared Spectroscopy

为了研究近红外光谱模型的优化方法,提高模型的精度,利用遗传算法对64个掺加了肉骨粉的鱼粉样品近红外光谱进行变量筛选,采用偏最小二乘法回归建模,并用21个样品进行外部验证。遗传算法共选取310个波长变量,相对于全谱的1556个变量减少了80%,与全谱范围的偏最小二乘法相比,交互验证相关系数(RCV)从0.80提高到0.90,交互验证均方根误差从5.22%降低到3.62%,预测相关系数(RV)从0.91提高到0.96,预测均方根误差从3.85%降低到2.95%,模型的稳健性和预测精度都显著提高。试验结果表明遗传算法可以改善近红外光谱法预测鱼粉中肉骨粉含量的效果。

For the purpose of optimizing near infrared spectroscopy model,and improving the prediction result,Genetic Algorithm （GA） was used to select wavelength variables of near infrared spectroscopy for fishmeal adulterated with meat and bone meal. 310 wavelengths are selted in genetic algorithm. By contrast with all wavelengths based partial least squares（PLS）,GA based PLS reduced 80% of the wavelengths,and gained much better cross validation and prediction results. Related coefficient of cross-validation RCV was improved from 0.80 to 0.90,while the value of root mean square error of cross- validation （RMSECV） was reduced from 5.22% to 3.62%. The related coefficient of prediction RV was improved from 0.91 to 0.96,while the value of root mean square error of prediction （RMSEP） was reduced from 3.85% to 2.95%. GA improved the robustness and predictability of the model. It’s indicated that GA was an effective method for variable selection and could improve the prediction result of the meat and bone meal content in fishmeal by near infrared spectroscopy.