Research On Near Infrared Spectroscopy Modeling Based On Quaternion Convolution Neural Network

To address the problems of weak absorption signals,overlapping spectral peaks,background and noise interference in Near Infrared Spectroscopy(NIRS),the establishment of high-performance qualitative and quantitative models with effective chemometric algorithms is one of the research hotspots in NIRS analysis technology.Convolutional Neural Network has the characteristics of local connectivity and weight sharing,which can efficiently process large amount of spectral data.In this paper,we propose a new method of serial fusion and quaternion parallel fusion spectroscopy combined with CNN to establish a quantitative model for maize composition content prediction and a qualitative and quantitative model for starch species identification and composition prediction.The main research contents are as follows:Firstly,in view of the international open corn near infrared spectrum data set,the sample original spectrum,first derivative and second derivative spectrum spectrum serial fusion,establishing corn sample moisture,fat,protein and starch in four kinds of quantitative prediction model of CNN,and merge with serial spectra combined with traditional regression algorithm respectively,comparing the PLS and SVR model,the results show that the model built by serial fusion spectrum combined with CNN has the best performance;Secondly,the experiment collects the near-infrared spectra of five types of starches of potato,corn,wheat,cassava and peas.Based on the quaternion algebra and convolutional neural network theory,five types of starch near-infrared spectroscopy quaternion convolutional neural network type identification models are established.The research results show that the performance of the classification model based on QCNN is significantly better than the performance of the classification model based on KNN,SVM and CNN combined with the optimal preprocessing algorithm;Finally,the experiment configures mixed samples of wheat and corn starch with different percentages.The near-infrared spectrum of mixed starch added with 3m L,4m L and 5m L of distilled water was embedded in the quaternion space and expressed as a pure quaternion spectrum matrix;a QCNN quantitative prediction model of mixed starch near-infrared spectroscopy was established.The research results show that compared with the quantitative model built by traditional regression algorithms PLSR and SVR,the model built by QCNN has better performance indicators and higher prediction accuracy.In this paper,serial fusion spectroscopy and quaternion parallel fusion spectroscopy are combined with convolutional neural network to strengthen the connection between spectral data and fully mine spectral information without preprocessing,which provides a new method and a new idea for the near-infrared spectroscopy analysis technology to more easily and quickly establish high-performance qualitative and quantitative models.