为了探索基于介电谱无损检测库尔勒香梨糖度和硬度的可行性，本文采用同轴探头和矢量网络分析仪无损测量了三个来源共168个库尔勒香梨在20~4500 MHz范围内201个频率点下的相对介电常数和介质损耗因子，分析了介电参数的变化规律。基于x-y 共生距离的样本划分法分别将112和56个样品划分为校正集与预测集；采用无信息变量消除法和连续投影算法（SPA）分别从全介电谱 (FS) 中提取出预测糖度的144个和20个特征变量，以及预测硬度的52个和9个特征变量；基于FS和提取的特征变量建立了预测库尔勒香梨糖度和硬度的偏最小二乘、误差反向传播、最小二乘支持向量机和极限学习机（ELM）模型。结果表明：SPA-ELM模型对库尔勒香梨糖度的预测效果最好，其校正相关系数（Rc）和预测相关系数（Rp）分别为0.943和0.876，校正均方根误差和预测均方根误差分别为0.298°Brix和0.330°Brix。所有硬度预测模型的Rc和Rp均小于0.6。研究结果说明，介电谱可用于无损检测库尔勒香梨的糖度，但对硬度的预测效果较差。后期仍需进一步研究介电谱无损检测库尔勒香梨硬度的可行性。

Dielectric spectra of Korla fragrant pears were used for nondestructive measurement of their sugar content and firmness. For this purpose, the dielectric constant and dielectric loss factor at 201 discrete frequencies across the frequency range of 10 to 4500 MHz were measured in 168 Korla fragrant pears from different areas, by means of an open-ended coaxial-line probe and vector network analyzer. Patterns of changes in the dielectric properties were analyzed. Sample-set partitioning based on the method of joint x-y distances was used to group 112 samples as a calibration set and 56 samples as a prediction set. To the full dielectric spectra (FS), uninformative-variable elimination and successive projection algorithm (SPA) were applied in order to extract 144 and 20 characteristic variables (CVs) for prediction of sugar content, and to extract 52 and 9 CVs for firmness prediction, respectively. The FS and extracted CVs were subjected to partial least squares, back propagation network, least square support vector machine, and extreme learning machine (ELM) analyses to establish sugar content and firmness measurement models for Korla fragrant pears. The results revealed that the SPA-ELM model showed the best prediction for sugar content of the Korla fragrant pears: correlation coefficients of the calibration and prediction sets were 0.943 and 0.876, respectively, and the root-mean-squared errors of the calibration and prediction sets were 0.298°Brix and 0.330°Brix, respectively. The correlation coefficients of the calibration and prediction sets of all predicted models for firmness were less than 0.6. The study indicates that dielectric spectra can be used to nondestructively measure sugar content of Korla fragrant pears. However, this approach shows poor predication performance on firmness. Further studies should be conducted to evaluate feasibility of the use of dielectric spectra to nondestructively measure the firmness of Korla fragrant pears.

国家自然科学基金资助项目（31171720）