Extreme Learning Model And Application Research On Nondestructive Testing Of Watermelon Ripeness

Ripeness is a key index for evaluating the quality of watermelons.A number of approaches have been proposed for nondestructive quality determination of watermelons,such as near infrared methods,electrical and magnetic methods,X-ray methods,acoustic methods etc.However,it remains an open challenge to develop easily implementable,low cost,high accuracy,intelligent methods of evaluating watermelon quality.This paper focuses on the acoustic detection methods of watermelon ripeness aiming at three maturity levels: unripe,ripe and overripe.A new method,based on the principal component analysis(PCA)and kernel principal component analysis(KPCA),is proposed to extract features from the acoustic signals of watermelons.And the extreme learning method(ELM)is first introduced to watermelons ripeness classification and sugar content detection.The main tasks are as follows:1.For acoustic signal of watermelons,a feature extraction method based on(kernel)PCA is proposed.The excited principal components is obtained from acoustic signals of different ripeness levels using(kernel)PCA,and those corresponding eigenvectors is linearly expanded into a finite dimensional feature space.The orthogonal projection coefficients of signals in feature space are regarded as features.PCA can reduce the dimensionality while preserving enough original information,and its calculation is straightforward and simple.The dimension of original signal can be reduced from 4096 to 31 and to 90 by PCA and KPCA respectively,when preserving 95% principal component.2.Extreme learning machine based on Markov chain sampling is proposed.ELM is a single hidden layer feedforward neural network.Its input layer weights and hidden layer threshold are randomly generated.The core of ELM is to calculate the Moore-Penrose generalized inverse of the output matrix.Hence,ELM is simple and fast in running,and performs better than some machine learning algorithms,such as support vector machine(SVM),BP neural network.Based on the theory of Tikhonov regularization,the generalization analysis of ELM under independent and identically distributed(i.i.d)samples is extended.And the upper bound of the misclassification error for ELM with Markov chain sampling is estimated.Empirical evaluations on real-word datasets are provided to compare the predictive performance of ELM with i.i.d and Markov chain sampling.The results demonstrate that the Markov chain sampling can not only effectively reduce the prediction error,but also improve the robustness of ELM.3.The KPCA-ELM based on Gaussian kernel is constructed as classification of watermelon ripeness.Firstly,we divide the samples with the ratio of 2:1,thus obtain 180 training samples and 90 testing samples.We analyze the effects of features extracted by PCA and KPCA on the classification of watermelon ripeness using ELM.Meanwhile,the effects of linear,polynomial,Gauss and Sigmoid kernel functions on the performance of KPCA-ELM are discussed.Finally,we compare ELM with K-nearest neighbor(KNN),BP neural network and SVM on classification accuracy and efficiency.The experimental results show that the KPCA-ELM based on Gaussian kernel performs the best.More specifically,the accuracy of watermelon ripeness detection on binary and three classifications can reach 95.72% and 89.23%,respectively.4.The KPCA-ELM based on Gaussian kernel is constructed for the detection of watermelon sugar content.Firstly,we use ELM to construct the regression model between the sugar content and acoustic signal,and analyze the effects of features extracted by PCA and KPCA on the detection of watermelon sugar content.In addition,we compare ELM with partial least squares(PLS),BP neural network,support vector regression(SVR).The experimental results show that KPCA-ELM achieves the best performance.More particularly,it obtains the minimum root mean square error(RMSE)as 0.3725 and the standard deviation(STD)as 0.0173.