基于介电频谱的灵武长枣可溶性固形物含量的预测模型

为了探寻利用介电频谱预测灵武长枣可溶性固形物含量的可行性,并建立最优模型,该文采用网络分析仪在200 MHz~18 GHz的频率范围内,选取101个频率点,测定分析了300个灵武长枣的介电损耗因子ε″和介电常数ε'频谱,利用长枣的介电损耗因子ε″和介电常数ε'进行了可溶性固形物含量的预测模型研究。通过遗传算法(genetic algorithm,GA)和相关系数法(correlation coefficient,CC)提取了介电频谱的有效信息,并分别采用偏最小二乘(partial least squares,PLS)、主成分回归(principal component regression,PCR)和支持向量机(support vector machines,SVM)法比较建立了可溶性固形物含量的预测模型。研究结果表明:用GA与CC方法提取频谱有效信息的建模效果要优于原始频谱的建模效果;PCR法的建模效果要优于PLS与SVM法的建模效果。以介电损耗因子ε″、介电常数ε'频谱建立的可溶性固形物含量的最优预测模型分别为GA-PCR和CC-PCR。以介电损耗因子ε″建立的GA-PCR模型优于介电常数ε'的CC-PCR模型,其校正集和预测集的相关系数分别为0.933和0.925,均方根误差分别为0.661%和0.702%。结果表明,利用介电频谱预测灵武长枣的可溶性固形物含量是可行的。

LingWu long jujube soluble solids content predicting model research based on dielectric spectra

Lingwu long jujube, as one of the special advantage fruits inNingxia Hui Autonomous Region, is favored by consumer. The traditional methods used in fruit soluble solids content detection are destructive testing and can’t satisfy the fruit commercialization testing requirements because the fruit has lost its commodity value after testing. With simple principle and strong adaptability, dielectric spectrum detection technology is easy to operate and nondestructive, and has become the development trend of fruit quality nondestructive testing in recent years. In order to explore the possibility of predicting LingWu long jujube soluble solids content and to establish optimal prediction model of long jujube soluble solids content based on the dielectric spectrum, dielectric loss factorε″ spectra and dielectric constantε′ spectra of 300 long jujube were measured with a network analyzer under 101 selected frequency points in the frequency range of 200 MHz-18 GHz. The prediction model of soluble solids content was researched using the long jujube dielectric loss factorε″ spectra and dielectric constantε′ spectra. The 300 jujubes are divided into calibration set and prediction set according to the proportion of 4:1 using K-S method. The influences of frequencies and storage time on long jujube dielectric parameters were discussed with variance analysis method. The effective information of the dielectric spectra was extracted by genetic algorithm (GA) and correlation coefficient method (CC). Prediction model of soluble solids content was established using partial least squares (PLS), principal components regression (PCR) and support vector machine (SVM). The best modeling method was acquired by comparative analysis of determination coefficientR2, the standard deviation RMSEC and the standard deviation RMSEP. The results indicated that, with the frequency increasing, dielectric loss factorε″ of long jujube decreased first and then increased, while dielectric constantε′ decreased gradually. The polarization characteristic frequency ?r of long jujube in the GHz frequency segments was 5.74 GHz gained from the spectrum curve. The analysis of variance indicated that frequency and storage time had a significant influence on dielectric properties. It is feasible to predict the soluble solids content based on the dielectric spectra. In order to improve the reliability of sugar prediction model, the effective information extraction methods of GA and CC were analyzed comparatively. 15 characteristic frequency points of the dielectric loss factorε″ and 14 characteristic frequency points of the dielectric constantε′ were optimized by GA. 14 effective frequency points of the dielectric loss factorε″ and 19 effective frequency points of the dielectric constantε′ were optimized by CC. Among them, the polarization characteristic frequency ?r (5.74 GHz) ofε″ andε′ could be found by both GA and CC. The modeling effect with the extracted effective information using GA and CC is better than that with original spectrum modeling, the modeling effect with PCR is better than that with PLS and SVM. The optimal prediction models of soluble solids content based on dielectric loss factorε″ and dielectric constantε′ spectra were GA-PCR and CC-PCR respectively. The effect of GA-PCR modeling with dielectric loss factorε″ is better than that of CC-PCR modeling with dielectric constantε′, the correlation coefficients of calibration set and prediction set are 0.933 and 0.925 respectively, and the root mean square errors (RMSE) were 0.661 and 0.702.