Study On Detection Method Of Rice Varieties And Adulteration Based On Machine Learning

Rice is the main grain variety in China,and its security is directly related to social stability and economic development.Due to the variety of rice and different storage conditions,especially in the circulation field,the quality of rice varies greatly,such as quality deterioration and adulteration.Therefore,timely detection becomes necessary to maintain the safety of rice.At present,there are mainly artificial methods,biochemical techniques,spectrum,wave spectrum,and mass spectrometric analysis for rice detection.Manual methods rely on experience,convenient and quick,but it is difficult to be effective in quantitative detection and simultaneous judgment of multiple results;Biochemical technology is based on the study of the components of the cell,namely protein,carbohydrate,lipid,nucleic acid,using the method of chemical synthesis to solve the facing problems,and in the detection of rice the sample preparation process is cumbersome and the technical requirements are high;The spectrum,wave spectrum and mass spectrometry were corresponding to terahertz time-domain spectroscopy,Raman spectroscopy and gas chromatography-ion mobility spectrometry respectively.The detection data of the same10 kinds of rice were obtained by using the three methods;the data of 5 kinds of rice with different contents of Bt63 transgenic components were obtained by terahertz time-domain spectroscopy;the rice data of Wuchang Daohuaxiang were also obtained by using gas chromatography-ion mobility spectrometry,which was used to identify adulteration.Combined with machine learning methods,such as multi-classification maximum-margin twin support vector machine and semi-supervised generative adversarial network,and through comparative study,the optimal identification method is optimized to achieve accurate identification of rice variety classification and adulteration.The methods and results are as follows:Terahertz time domain spectroscopy and Raman spectroscopy were used to collect the data of different rice varieties,and support vector machine was used to classify and identify the spectral data.In order to reduce the complexity of support vector machine model,principal component analysis is used to reduce the dimension of sample space,which improves the classification performance of support vector machine.Aiming at the improvement of support vector machine,a multi-classification maximum-margin twin support vector machine is proposed to extend the binary classification to multi-classification,which improves the running efficiency of support vector machine in multi-classification.The improved support vector machine was used to identify rice varieties,and the recognition rate based on terahertz time-domain spectroscopy,Raman spectroscopy reached 91.60%,94.23% respectively.The gas chromatography-ion mobility spectrometry of different rice varieties and adulterated samples was collected and calculated,the characteristic spectrum of ion mobility spectrometry formed.Considering the strong recognition ability of convolutional neural network for image and the problem of deep learning method needing of a large number of training samples,the number of samples was amplified and the type of samples was identified by generative adversarial network.In order to improve the identification accuracy of generative adversarial network for samples,a semi-supervised generative adversarial network method was proposed.By introducing semi supervised mode into the traditional generative adversarial network and a small number of labeled samples can be used for auxiliary training,which reduces the cost of model training,optimizes the model parameters and improves the model classification accuracy.The improved generative adversarial network was used to identify rice varieties,and the recognition rate reached 98.00% based on gas chromatography-ion mobility spectrometry;the identification rate of rice adulteration reached 97.30%.The results show that the recognition performance of the generative adversarial network is better than that of the improved support vector machine for the sample data from gas chromatography-ion mobility spectrometry.However,the algorithm complexity of support vector machine is obviously lower than that of generative adversarial network.For the data from terahertz time-domain spectroscopy,the improved support vector machine method is not superior to the traditional support vector machine in classification performance,but the time complexity of the algorithm is significantly reduced.At the same time,compared with other detection methods,the method of gas chromatography-ion mobility spectrometry combined with semi-supervised generative adversarial network,which is used in this paper,has obvious improvement in both detection accuracy and convenience.