Quantitative analysis of melamine in terahertz spectra based on convolutional neural network
[Objective]Combining convolutional neural networks and terahertz time-domain spectroscopy for quantitative analysis of the illegal additive melamine in milk powder.[Methods]The terahertz absorption spectra of melamine,milk powder both individually and in mixtures were measured using a transmission-type terahertz time-domain spectroscopy system.Various methods such as Savitzky-Golay(S-G)smoothing,Gaussian smoothing,moving average,and R-Loess smoothing were employed to correct the original spectral data.A partial least squares(PLS)regression model was established,and the optimal terahertz spectroscopic correction preprocessing method was determined by comparing model evaluation criteria.The PLS model corrected with S-G smoothing was chosen for the quantitative analysis of the mixed samples.Quantitative regression models based on partial least squares(PLS),least squares support vector machine(LS-SVM),backpropagation neural network(BPNN),and convolutional neural network(CNN)were separately established,and the content of melamine in the mixed samples was predicted.[Results]The correlation coefficients of the prediction set for the PLS,LS-SVM,BPNN,and CNN models were 0.997 1,0.997 7,0.998 1,and 0.998 7,respectively,with prediction set root mean square errors of 0.551%,0.494%,0.437%,and 0.374%,respectively.[Conclusion]Compared to the other three models,the CNN regression model has the highest prediction accuracy and is more suitable for accurately detecting the content of melamine in milk powder.
terahertz time domain spectroscopyquantitative analysisspectral correctionconvolution neural network
国家科技期刊平台
NSTL
万方数据
