Application And Research Of Terahertz Spectroscopy In The Field Of Pesticide Quantitative Detection

In modern agricultural production,pesticides play an extremely important role.However,with the large-scale use of pesticides,some negative effects are gradually exposed.Among them,pesticide residues in agricultural products have attracted consumers' attention.Realizing the quantitative detection of pesticide residues in agricultural products is a guarantee for consumers' safe consumption.Traditional detection methods have some shortcomings in detection time,detection accuracy,sample destructiveness,etc.,while terahertz spectroscopy technology has the characteristics of fast,accurate,and non-destructive,which shows a broad application prospect in the qualitative and quantitative detection of pesticides.In this paper,three pesticides,6-Benzylaminopurine(6-BA),imidacloprid,2,6-Dichlorobenzonitrile(2,6-D)and agricultural product wheat flour were selected as the detection objects.Using terahertz spectroscopy technology,combined with machine learning algorithms,I conducted research from two aspects: the quantitative detection of pesticides in agricultural products and the selection of high-efficiency spectrum bands.The work of this paper mainly included the following three parts:(1)Quantitative detection of a single pesticide.We mixed 6-benzylaminopurine and polyethylene in different proportions at high concentrations,and chose the partial least squares regression(PLS)to establish a quantitative model of absorption spectra and pesticide concentration.Then,we selected root mean square error(RMSE)and the correlation coefficient(R)to evaluate the performance of the model.This work proves the feasibility of combining machine learning algorithms and terahertz spectroscopy to achieve quantitative pesticide detection.(2)Quantitative detection of low-concentration pesticide mixtures in actual agricultural products.Three pesticides were mixed with wheat flour in different proportions at low concentrations,and the applicability of the three algorithms,PLS,BP neural network(BPNN),and support vector machine regression(SVR),was compared.The results showed that in the quantitative detection of pesticide multi-component mixtures,the two types of nonlinear models,BPNN and SVR,perform better than the linear model PLS,especially the BPNN with multi-parameter adjustment.In addition,in order to further improve the detection accuracy,the important parameters of BPNN(such as the number of neural network layers,the number of hidden layer nodes,the transfer function,the learning rate,etc.)of the BPNN have been optimized in combination with the genetic algorithm(GA).After optimization,the model accuracy has been significantly improved,the average correlation coefficient R was increased from 0.9323 to 0.9963.The optimized model had more practical reference and application value for realizing the quantitative detection of pesticides in agricultural products.(3)Selection of high-efficiency frequency bands of the spectrum.In addition to the selection of models,the selection of frequency bands of the spectrum is also important for detection accuracy.The original spectrum obtained in the experiment contains both high signal-to-noise ratio information and various types of noise,such as scattering effects and random interference.In this paper,the equal interval partial least square regression algorithm(i PLS)was used to detect the pesticide content in different intervals,and we found that the high-efficiency frequency band of the absorption spectrum is closely related to the absorption peak.In view of the particularity of the absorption peak,by introducing two parameters,the sum of peak interval(Si)and peak width,the selection of absorption peaks and the study of absorption peak width were realized,and the high-efficiency spectrum interval of the sample was determined.Comparing the quantitative analysis errors of selecting the original spectrum and the high-efficiency frequency band spectrum,the results showed that under the same model,the appropriate frequency band selection can significantly improve the quantitative detection accuracy.