基于介电特征的苹果霉心病检测方法

针对苹果霉心病无法有效根据外表进行识别，且传统检测方法具有设备复杂、成本高昂等问题，本研究通过采集苹果介电参数构建苹果霉心病检测模型，从而实现简单快速的苹果霉心病无损检测。基于LCR测量仪采集220个苹果的108项介电指标（9个频率下的12项介电指标）作为原始参数，使用数据标准化、主成分分析算法等对数据进行预处理，并利用BP神经网络、支持向量机、随机森林算法构建霉心病果检测模型。试验结果表明，基于随机森林算法构建的霉心病果检测模型性能最佳，在150个苹果构建的训练集和70个苹果构建的测试集中分类准确率分别达到96.66%和95.71%；基于采用BP神经网络构建的霉心病果检测模型效果次之，分类准确率分别可达到94.66%和94.29%；基于使用支持向量机构建的模型检测效果相对较差，分类准确率分别为93.33%和91.43%。试验结果表明，使用随机森林构建的模型可以更有效地识别霉心病果和好果。本研究可为苹果病虫害及苹果品质无损检测等提供参考。

Detection Method of Apple Mould Core Based on Dielectric Characteristics

Apple mouldy core disease often occurs in the ventricle of apples and cannot be effectively identified by appearance. Near-infrared spectroscopy, nuclear magnetic resonance and other methods are usually used in traditional apple mouldy core disease detection, but these methods require complex equipment and high detection costs. In this research, a simple and fast nondestructive detection method of apple mouldy core disease was proposed by using a dielectric method to construct an apple mouldy core disease detection model. Japan's Hioki 3532-50 LCR tester was used to collect 108 dielectric indicators (12 dielectric indicators at 9 frequencies) of 220 apples as the original data. Due to the large differences in the distribution of data collected with different dielectric indexes and different frequencies, a standardized method was used for data preprocessing to eliminate the problem of large differences in dielectric data distribution. Afterwards, in order to eliminate the redundant information between the data, the principal component analysis algorithm was used to reduce the data dimensionality, and finally the three algorithms of BP neural network (BPNN), support vector machine (SVM) and random forest (RF) were used to construct the mouldy core disease detection model. After pre-experiment, the most effective parameters of each algorithm were selected, the test results showed that the apple mouldy core disease detection model based on the RF algorithm obtained the best performance, and the detection accuracy rate reached 96.66% and 95.71% in the training set (150 apples) and the test set (70 apples). The mouldy core disease detection model constructed by using BPNN was the second most effective, and the detection accuracy could reach 94.66% and 94.29%, respectively. The detection effect of the model built by using SVM was relatively poor, and the detection accuracies were 93.33% and 91.43%, respectively. The experimental results showed that the model constructed by using RF can more effectively identify mouldy core disease apples and healthy apples. This study could provide references for apple diseases and insect pests and non-destructive testing of apple quality.