霉变玉米中玉米赤霉烯酮和黄曲霉毒素B1含量电子鼻检测模型的构建

为建立霉变玉米中玉米赤霉烯酮和黄曲霉毒素B1的电子鼻检测方法，首先以电子鼻对7 级不同霉变程度玉米响应信号的积分值作为特征参量，然后分别利用主成分回归、偏最小二乘回归、BP（back-propagation）神经网络、最小二乘支持向量机等方法建立霉变玉米中玉米赤霉烯酮与黄曲霉毒素B1含量的预测模型，并进行了比较分析。结果表明，主成分回归预测精度最差，偏最小二乘回归较差、BP神经网络和最小二乘支持向量机法比较好。对于玉米赤霉烯酮，4 种预测模型70 个样本中相对误差控制在5%以内的样本数分别为23、45、63、67 个。对于黄曲霉毒素B1，4 种预测模型70 个样本中相对误差控制在5%以内的样本数分别为19、41、62、65 个。同时，变换不同的训练集和测试集以考察BP神经网络、最小二乘支持向量机建模方法的稳健性，结果表明，在BP神经网络结构和最小二乘支持向量机核函数与核函数参数均未发生改变的条件下，两种建模方法依然有较高的预测精度，这说明了两种模型具有较高的稳健性。

Predictive Models for the Detection of Zearalenone and Aflatoxin B1 Contents in Moldy Corn with Electronic Nose

This study aimed to explore a quantitative method for detecting the contents of zearalenone and aflatoxin B1 in
moldy corn using electronic nose. Firstly, the integral values of the electronic nose response signals of corn samples with
different mildew levels were extracted and used as feature parameters for establishing a predictive model for predicting
the contents of zearalenone and aflatoxin B1 in moldy corn samples employing principal component regression (PCR),
partial least squares regression (PLSR), back-propagation (BP) neural network, and least squares support vector machine
(LS-SVM), respectively. The results from the different models developed were compared. It was shown that the prediction
accuracy of the PCR model was the worst among four models, the PLSR model had better prediction accuracy, and the
BP neural network and LS-SVM models provided the most accurate predictions. The PCR, PLSR, BP neural network and
LS-SVM models gave good predictions of zearalenone with relative errors less than 5% for 23, 45, 63, and 67 out of 70
samples, respectively, while they provided good predictions of aflatoxin B1 with relative errors less than 5% for 19, 41, 62
and 65 out of the 70 samples. At the same time, different training and test sets were used to examine the robustness of the BP
neural network and LS-SVM models. The results showed that the BP neural network architecture, and the kernel function
and kernel parameter of LS-SVM remained unchanged. The prediction accuracy of the two models was still good, showing
that both models are of high robustness.how