【作者】 程绍明；

【导师】 刘建新； 王俊；

【作者基本信息】 浙江大学 ， 农业生物环境与能源工程， 2011， 博士

【摘要】 植物种子会产生挥发物,这种挥发物与种子活性相关。植物体在受到损伤和病虫的危害后会产生特异性挥发物,使用电子鼻技术可以检测这种挥发物的变化情况,进而对损伤情况、病害情况进行诊断。本研究采用电子鼻技术对不同发芽率的番茄种子、不同损伤处理和不同病害处理的番茄苗样本进行了检测,应用数据特征选择与提取、模式识别等方法对检测的数据进行分析,研究不同发芽率的番茄种子、不同损伤番茄苗和不同病害番茄苗的挥发物与电子鼻响应信号的关系,建立了各种模型,验证了使用电子鼻技术对不同发芽率的番茄种子、不同损伤番茄苗和不同病害番茄苗检测的可行性。得到的主要结果如下：(1)番茄种子发芽率的检测研究利用电子鼻系统对不同发芽率(发芽率分别为0%、50%、60%、70、80和90%)的番茄种子进行检测,通过主成分分析、线性判别分析对六种不同发芽率的番茄种子进行研究,表明采用PCA分析、LDA分析可快速的区分出番茄种子发芽率为0%、70%、80%和90%的四种情况；当番茄种子发芽率为50%、60%和70%时,其图形信息部分重叠,表明利用电子鼻较难将这三种情况区分开。采用BP神经网络时训练集和预测集准确率分别为93.6%、65.2%；采用支持向量机时训练集和预测集准确率分别为97.44%、74.73%。对比BP神经网络和支持向量机分析结果,支持向量机的预测结果要好一些。(2)番茄苗机械损伤的检测研究番茄苗受到机械损伤会引起电子鼻传感器信号的变化,说明番茄苗受到机械损伤后其释放出的挥发物发生了改变。利用电子鼻对不同处理(对照组、30针、60针和90针)的机械损伤番茄苗进行检测,采用主成分分析、线性判别分析对四种不同处理机械损伤的番茄苗进行研究,结果表明主成分分析各处理样本间均有重叠,区分效果不理想,线性判别分析各处理样本基本可以分开,表明使用线性判别分析的效果比主成分分析效果要好。用逐步判别分析和BP神经网络分析对四种不同处理样本进行模式识别,结果表明,采用逐步判别分析和BP神经网络分析时的测试集的准确率分别为84.4%和90%,BP神经网络模型相对于逐步判别分析的预测结果更好。(3)番茄苗早疫病和灰霉病病害的检测研究对35d苗龄的番茄苗进行早疫病、灰霉病病菌接种,按每株接种1叶片、2叶片、4叶片和对照四个处理进行,然后在高温、高湿环境中培养,使病菌快速入侵叶片。将培养后的番茄苗利用电子鼻和顶空采样装置对感染病害的番茄苗进行了检测。利用主成分分析方法可以观察出番茄苗受害程度变化的规律,但是样本之间有很多重叠现象；线性判别分析法可以区分出受到病害与没有受到病害的番茄苗样本,不同受害程度的样本只有个别重叠；表明使用线性判别分析的效果比主成分分析效果要好。将四组不同受害程度的番茄苗每组随机抽取10个样本作为训练集(共40个样本),用于对番茄苗病害程度的训练,剩余每组6个样本(共24个样本)作为预测集。使用逐步判别分析对番茄苗早疫病病害程度进行识别,番茄苗受到早疫病病害24h后的训练集交叉验证的正确率为100%,预测的平均正确率为50%。番茄苗受到早疫病病害48h后的训练集交叉验证的正确率为97.5%,预测的平均正确率为66.7%。使用BP神经网络分析对早疫病病害程度进行识别,番茄苗受到早疫病病害24h后的训练集交叉验证的正确率为95%,预测的平均正确率为87.5%。番茄苗受到病害48h后的训练集交叉验证的正确率为92.5%,预测的平均正确率为79.2%。使用逐步判别分析番茄苗灰霉病病害程度进行识别,番茄苗受到灰霉病病害24h后的训练集交叉验证的正确率为97.5%,预测的平均正确率为70.8%。番茄苗受到病害48h后的训练集交叉验证的正确率为82.5%,预测的平均正确率为62.5%。使用BP神经网络分析对灰霉病病病害程度进行识别,番茄苗受到灰霉病病害24h后的训练集交叉验证的正确率为92.5%,预测的平均正确率为87.5%。番茄苗受到病害48h后的训练集交叉验证的正确率为82.5%,预测的平均正确率为66.7%。(4)不同特征值对番茄苗病害识别效果的研究采用不同特征值(最大值、全段数据平均值、相应曲线的全段积分值和响应曲线最大曲率)作为电子鼻对感染早疫病和灰霉病病害24h和48h的番茄苗样本的响应信号进行PCA和LDA分析,并采用遗传神经网络与BP神经网络模型进行识别分析。对比不同特征选择方法的预测结果,全段数据平均值(Mean)和响应曲线的全段积分值(IV)的训练集和预测集的正确率较好,其次为最大值(Max)方法,预测结果最差的是响应曲线最大曲率(Kmax)方法。将不同病害、不同病害时间的番茄苗分别每组随机抽取10个样本作为训练集(共40个样本),用于对番茄苗病害程度的训练,剩余每组6个样本(共24个样本)作为预测集。对比BP神经网络和GABP神经网络两种模型结果,GABP神经网络分析的决定系数均要高于BP神经网络,说明经过遗传算法优化后的BP神经网络模型的预测结果要好于BP神经网络。(5)不同种类损伤番茄苗的检测研究通过电子鼻响应信号的对比图,机械损伤、早疫病病害、灰霉病病害和对照组四个不同处理的番茄苗样本电子鼻的响应信号是不同的。采用主成分分析和线性判别分析法对各处理番茄苗样本的电子鼻数据进行特征提取和降维分析,三维主成分图和线性判别分析图可以区分不同处理的番茄苗样本。采用聚类分析方法得到与PCA相类似的结果,不同处理番茄苗样本可以用聚类分析方法进行区分。选择40个番茄苗样本作为训练(每组10个样本),24个样本作为测试组(每组6个样本)。采用BP神经网络和支持向量机模型对各番茄苗处理样本进行预测,两种模型都取得了较好的预测结果,预测平均正确率均为83.3%。说明电子鼻技术可以用于对番茄苗样本的实际检测。综上所述,我们可以得到如下结论：(1)利用电子鼻技术能对不同发芽率的番茄种子进行区分；(2)利用电子鼻技术能识别感染不同病害和不同程度的番茄苗；(3)利用电子鼻能区分不同种类损伤的番茄苗；更多还原

【Abstract】 Plants alter their profiles of emitted volatiles in response to damage or herbivore attack. This study investigated the potential of the electronic nose technology to monitor such changes, with the aim to diagnose plant health. In this study, electronic nose (E-nose) was used to analyse different treatment of tomato seeds and tomato plant samples. Several statistical methods such as feature extraction and pattern recognition were used for analyzing the experimental data. We have studied the relations between E-nose response signals and disease-induced volatiles, and developed the pattern recognition models. The results demonstrated that it is plausible to use E-nose technology as a method for monitoring damage in tomato cultivation practices. The main conclusions were as follows:(1) Research on detecting tomato seeds with different germinationAn investigation was made to evaluate the capacity of an electronic nose to classify the tomato seeds with different germination. The data were processed by Principal Component Analysis (PCA) and Linear Discrimination Analysis (LDA). The result shows that the electronic nose can distinguish the tomato seeds with germination percentage of 90%,80%,70%-50% and un-germination seeds. However, samples with germination percentage of 50%,60% and 70% were overlapped. The e-nose can distinguish the different time tomato seeds which blended each at different proportion by using BPNN and SVM. A better predicted rate was obtained by SVM than BPNN.(2) Research on detecting mechanical-damageThe value of E-nose response signals differed with different levels mechanical (Opricks,30pricks,60pricks and 90pricks) damaged tomato plants, indicating that the emission of volatiles by tomato plants changes in response to different degrees of damage. Stepwise discriminant analysis (SDA) and back-propagation neural network (BPNN) were applied to evaluate the data. The average correction ratios of testing set of SDA and BPNN were 84.4% and 90%. The results obtained indicate that it is possible to classify different degrees of damaged rice plants using e-nose signals.(3) Research on detecting Early blight disease and Gray mold disease infestation An E-nose equipped with a headspace sampling unit was used to discriminate tomato plants infested Early blight disease at different times. PCA resulted in an even distribution of the 4 different treatments with different disease densities. LDA was able to distinguish between all treatments. The discrimination rates were over 100% for training data sets and 50% for the testing sets, as determined using SDA in 24h later. The discrimination rates were over 97.5% for training data sets and 66.7% for the testing sets, as determined using SDA in 48h later. The discrimination rates were over 95% for training data sets and 87.5% for the testing sets, as determined using BPNN in 24h later. The discrimination rates were over 92.5% for training data sets and 79.2% for the testing sets, as determined using BPNN in 48h later. The results indicated that it is possible to predict the different diease densities in tomato plant using the E-nose.The discrimination rates were over 97.5% for training data sets and 70.8% for the testing sets, as determined using SDA in 24h later. The discrimination rates were over 82.5% for training data sets and 62.5% for the testing sets, as determined using SDA in 48h later. The discrimination rates were over 92.5% for training data sets and 87.5% for the testing sets, as determined using BPNN in 24h later. The discrimination rates were over 82.5% for training data sets and 66.7% for the testing sets, as determined using BPNN in 48h later. The results indicated that it is possible to predict the different disease densities in tomato plant using the E-nose.(4) Different characteristic value using for BPNN and GABPFour kinds of characteristic value (Mean,Kmax,IV and Max) were used to analyse the different infestation treatments of tomato paint. Back-propagation neural network (BPNN) and genetic algorithms back propagation network (GABP) were developed for pattern recognition models. When the models were used to predict infestation degrees using different feature selection methods. The highest coefficient of determination between predicted and real numbers of the disease using the IV method. Compare BP results with GABP results, the coefficient of determinations of the GABP were higher than the BP. The results demonstrated that after optimized the BP neural network with genetic algorithm better results were obtained.(5) Research on detecting different types of damageFrom the chart, the E-nose sensor response changed for tomato plants by 4 types of treatments(ZP, HP, CP and MP). The results of PC A and LDA showed that clusters of data were divided into 3 groups (ZP, HP, and CP/MP). Samples from groups CP and MP overlapped partially. Similar results could be obtained when using CA. The results of the CA showed obvious differentiation between the tomato plant samples with different types of damage. Back-propagation neural network (BPNN) and support vector machine (SVM) network were used to evaluate the E-nose data. Good discrimination results were obtained using SVM and BPNN. The results demonstrate that it is plausible to use E-nose technology as a method for monitoring damage in tomato plant.更多还原