Identification And Classification Of Damaged Corn Particles Based On Collision Acoustic Signals

The objects of this paper are three types of corn,which are undamaged kernels,insect-damaged kernels and mildew-damaged kernels.The aim of the study is the identification and classification of damaged corn kernels.By analyzing the time-frequency characteristics of the three impact acoustic signals,the methods of feature extraction which are ensemble empirical mode decomposition,integration of multi-domain and dual space are used.Subsequently,these features are used as inputs to a support vector machine which is optimized by particle swarm optimization.With the experiments,good recognition effect are achieved.The main contents of this paper includes the following sections:Introduce the background and significance of this research and the application of impact acoustic signal in detection and identification of agricultural products,review of the domestic and foreign research.Describe experimental apparatus for collection of impact acoustic signal of corn kernels,experimental materials and feature analysis of the impact acoustic signal.Summarize the previous research methods of signal processing,find that features extracted by the traditional methods for feature extraction are relatively single.Thus it can not fully reflect the essential characteristics of the signal and result in lower separability of signals.Moreover,most of the features reflect the overall characteristics of the signal,while ignoring the local characteristics.Therefore,ensemble empirical mode decomposition and a multi-domain method are proposed.EEMD can self-adaptively process non-stationary signals,suppress mode mixing and analyze localcharacteristics of the signals.Time domain,frequency domain,and Hilbert domain features which are extracted from the impact acoustic signals can avoid simplification of features.Subsequently,these features are used as inputs to a support vector machine which is optimized by particle swarm optimization.This algorithm aims at finding the optimal parameters so as to overcome the flaws of conventional methods of support vector machine model.The use of hybrid features enable higher classification accuracy than usage of features in each domain separately.In this study,the optimal classification accuracies are 98.6%,99.2%,and 99.3%for undamaged kernels,insect-damaged kernels and mildew-damaged kernels,respectively.A dual space feature extraction method is put forward based on principal component analysis(PCA)and kernel independent principal component analysis(KICA),the PCA-KICA method.Impact acoustic signal device is used to collect signals of undamaged corn kernels,insect-damaged kernels and mildew-damaged kernels.Experimental results show,single subspace can not get a better classification accuracy rate,but applying dual space feature extraction method can overcome the limits of single subspace.PCA-KICA method emerges the highest recognition rates,which are 95.00%,96.40%,97.80%for undamaged kernels,insect-damaged kernels and mildew-damaged kernels respectively.The method provides a basis for the detection and identification of damaged corn kernels and other agricultural products.