西瓜成熟度无损检测的极限学习模型及应用研究

发布时间：2018-04-28 16:30

本文选题：敲击响应信号 + 成熟度　；参考：《华中农业大学》2017年硕士论文

【摘要】：西瓜成熟度无损检测是西瓜品质检测研究中的一个核心问题,近红外、核磁共振、X射线、声学等众多的方法被应用到西瓜成熟度检测。目前西瓜品质智能检测技术尚未在实践中推广,主要原因是检测算法的有效性、快速性、稳定性和易实现性等未达到应用上的要求。本文考虑未熟、成熟和过熟三个成熟度等级的西瓜声学分类问题,通过主成分分析(Principal component analysis,PCA)和核主成分分析(kernel principal component analysis,KPCA)提取西瓜敲击响应信号的特征,利用极限学习机(Extreme Learning Machine,ELM)构建西瓜成熟度分类和西瓜含糖量检测模型,主要的工作如下:1.提出基于(核)主成分分析的西瓜敲击响应信号的特征提取方法。通过不同成熟水平声信号样本的(核)主成分分析获得感兴趣的主成分,将其对应的特征向量线性扩张成有限维的特征空间,用信号在特征空间上的正交投影系数作为特征。PCA能够在保留足够多原始信息的同时大幅度降维,计算直接且简单。在保留95%主成分的条件下,PCA和KPCA分别将原始信号维数由4096降低到31和90。2.提出基于Markov chain采样的极限学习机算法。ELM是一个单隐层前馈神经网络学习系统,其输入层权重和隐层阈值是随机固定的,算法的核心是计算输出矩阵的Moore-Penrose广义逆。该算法简单、运行速度快,且优于支持向量机(SVM)、BP神经网络等机器学习算法。本文在Tikhonov正则化理论分析基础上,将独立同分布数据下的ELM泛化性能的进行推广,给出了基于Markov chain采样的ELM算法误差上界的估计。同时,基于真实数据,将独立同分布与Markov chain采样条件下的ELM进行了对比分析。实验结果表明,Markov chain采样不仅能有效降低ELM的预测误差,且能够提高ELM的鲁棒性。3.构建基于高斯核函数的KPCA-ELM西瓜成熟度分类模型。首先,按2:1的比例将样本进行随机划分,获得训练集样本180个,测试集样本90个。其次,以ELM为分类模型,分析了由PCA和KPCA提取的特征对西瓜成熟度检测的影响。同时,讨论了基于线性、多项式、高斯、S型四种核函数的KPCA-ELM分类模型对西瓜成熟度的检测效果。最后,从分类准确率和速率两个角度,将ELM与K-最近邻(KNN)、BP神经网络和SVM三种分类模型进行了对比分析。实验结果表明,基于高斯核函数的KPCA-ELM的模型效果最优,在二分类和三分类两种西瓜成熟度检测场景下的识别准确率分别为95.72%和89.23%。4.构建基于高斯核函数的KPCA-ELM西瓜含糖量检测模型。利用ELM构建西瓜含糖量与敲击响应信号之间的回归模型,分析了PCA、KPCA提取的特征对西瓜含糖量检测模型的影响。同时,也将ELM与偏最小二乘(PLS)、BP神经网络、支持向量回归(SVR)进行了对比分析。实验结果表明,基于KPCA-ELM的西瓜含糖量检测模型的性能最佳,其能够得到最小均方根误差(Root mean square error,RMSE)为0.3725,标准偏差(Standard deviation,STD)为0.0173。
[Abstract]:Non-destructive testing of watermelon maturity is a core problem in watermelon quality testing. Many methods, such as near-infrared, nuclear magnetic resonance X-ray, acoustics and so on, have been applied to watermelon maturity detection. At present, the intelligent detection technology of watermelon quality has not been popularized in practice, the main reason is that the validity, rapidity, stability and realizability of the detection algorithm do not meet the requirements of application. In this paper, the acoustic classification of watermelon with immature, mature and overripe grades was considered. Principal component analysis (PCA) and kernel principal component analysis (KPA) were used to extract the characteristics of the knock-response signals of watermelon by principal component analysis (PCA) and kernel principal component analysis (KPA). Using extreme Learning machine (ELM) to construct the watermelon maturity classification and sugar content detection model, the main work is as follows: 1. A method for feature extraction of watermelon knock response signal based on kernel principal component analysis (PCA) is proposed. The principal components of interest are obtained by kernel principal component analysis (PCA) of different mature horizontal acoustic signal samples, and the corresponding eigenvector is linearly expanded into a finite dimensional feature space. Using the orthogonal projection coefficient of the signal in the feature space as the feature. PCA can greatly reduce the dimension while retaining enough original information. The calculation is direct and simple. Under the condition of keeping 95% principal component, the original signal dimension was reduced from 4096 to 31 and 90.2 by KPCA, respectively. A learning algorithm based on Markov chain sampling is proposed. Elm is a single hidden layer feedforward neural network learning system. Its input layer weight and hidden layer threshold are randomly fixed. The core of the algorithm is to calculate the Moore-Penrose generalized inverse of the output matrix. The algorithm is simple, fast and superior to the support vector machine (SVM) SVM / BP neural network. Based on the analysis of Tikhonov regularization theory, this paper generalizes the generalization of the ELM generalization performance under the independent same distribution data, and gives the estimation of the upper bound of the error of the ELM algorithm based on the Markov chain sampling. At the same time, based on the real data, the ELM under the condition of Markov chain sampling and independent same distribution are compared and analyzed. The experimental results show that chain sampling can not only effectively reduce the prediction error of ELM, but also improve the robustness of ELM. The maturity classification model of KPCA-ELM watermelon based on Gao Si kernel function was constructed. First, the samples are randomly divided according to 2:1 scale, and 180 samples of training set and 90 samples of test set are obtained. Secondly, using ELM as the classification model, the effects of the characteristics extracted by PCA and KPCA on watermelon maturity were analyzed. At the same time, the effect of KPCA-ELM classification model based on linear, polynomial and Gao Si S-type kernel function on watermelon maturity was discussed. Finally, from the perspective of classification accuracy and rate, three classification models, ELM, KNNNNNNBP neural network and SVM, are compared and analyzed. The experimental results show that the model of KPCA-ELM based on Gao Si kernel function is the best, and the recognition accuracy is 95.72% and 89.23.4in two kinds of watermelon maturity detection scenarios, respectively. The sugar content detection model of KPCA-ELM watermelon based on Gao Si kernel function was established. The regression model between sugar content and knock response signal of watermelon was constructed by ELM, and the influence of extraction characteristics of ELM on sugar content detection model of watermelon was analyzed. At the same time, the ELM is compared with the partial least squares BP neural network and the support vector regression (SVR). The experimental results show that the model based on KPCA-ELM has the best performance, the minimum root mean square error (RMSE) is 0.3725, and the standard deviation (STD) is 0.0173.
【学位授予单位】：华中农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;S651

【参考文献】