In recent years, the food safety hazards caused by the illegal use and abuse of food additives during their use have once again attracted much attention from society. The research and development of non-destructive, fast, accurate, and efficient qualitative and quantitative detection technologies and methods for food additives have become a research hotspot for scholars. At present, traditional detection methods for food additives include ion chromatography, liquid chromatography, liquid chromatography-mass spectrometry, gas chromatography, and molecular spectrometry. The traditional methods have obvious shortcomings, such as high equipment cost, long detection cycle, high cost, uneven detection accuracy, high requirements for the operation of technicians and the purity of organic solvents, and complex detection operations. Compared with traditional detection methods, the application of new technologies such as immune detection, biosensor, and spectral analysis has supplemented and improved old technologies, thereby promoting the development of food additive detection technology. The detection method based on terahertz spectroscopy technology is non-destructive, fast, and efficient, and has been widely applied in fields of food, medicine, and environmental detection in recent years.

Firstly, we construct melamine samples with concentration gradients and obtain experimental training and testing sets by a transmission terahertz time-domain spectroscopy system. For the high-dimensional spectral data obtained from the experiment, a Savitzky-Golay convolutional filter is adopted for preprocessing to reduce quantitative prediction errors. Secondly, based on the dimensional characteristics of spectral data, we build four regression prediction models including PCR, SVR, PLSR, and LSSVR for data analysis. The obtained experimental spectral data are compared in terms of the predictive performance after linear regression dimensionality reduction (PCR, PLSR) and nonlinear regression dimensionality enhancement (SVR, LSSVR), which are processed at opposite angles. The correlation coefficient R_P of the prediction set and the root mean square error of the prediction set (RMSEP) are employed as indicators for model performance evaluation. Finally, according to the optimal evaluation index, we find that the prediction effect of the LSSVR model is optimal. We leverage particle swarm optimization (PSO), genetic algorithm (GA), Cuckoo search algorithm (CS), and grey wolf optimization (GWO) to calculate the regularization parameter C in LSSVR and the kernel parameter after the determined kernel function is Gaussian kernel function γfor parameter optimization.

The filtering preprocessing operation for spectral data yields sound effect (Fig. 1). We employ four different regression models (PCR, PLSR, SVR, and LSSVR) to predict the melamine content in milk powder, and adopt the correlation coefficient of the prediction set and RMSEP as the model evaluation coefficients. After comparing the evaluation coefficients of the four models, it is determined that the minimum correlation coefficient of the linear model PCR's prediction set is 0.99715, the maximum RMSEP is 0.50%, and the nonlinear model LSSVR has the best prediction performance. Its prediction phase set relationship number R_P is 0.99838 and RMSEP is 0.41%, which indicates that the nonlinear model has better detection performance for terahertz spectral data (Fig. 3). On this basis, we utilize swarm intelligence algorithms (PSO, GA, CS, and GWO) whose performances are significantly better than those of traditional methods to optimize hyperparameter selection of LSSVR model respectively. The prediction accuracy of the model after optimization by the four algorithms has been improved. Among them, the evaluation coefficient of the GWO-LSSVR model is the best, with R_P of 0.99925 and RMSEP of 0.28% (Fig. 4).

Results show that nonlinear models can be better applied to the detection of food additives by terahertz technology. The optimized GWO-LSSVR model can improve the accuracy of regression models in predicting mixture concentration and the quantitative detection accuracy of melamine based on terahertz spectral data. Additionally, it can promote the application of terahertz spectral technology in food additive detection and provide new methods and ideas for the quantitative analysis of food additives. However, the predictive performance and stability of the model are also relative, and the issues of computational complexity and time consumption should also be considered important factors in algorithm selection.