Detection of Adulterants in Egg White Powder Using Near-Infrared Spectroscopy Based on an Improved One-Dimensional Convolutional Neural Network

In response to the market regulation requirements for detecting adulterated egg white powder, based on the near-infrared spectroscopy (NIRS) data of pure and adulterated egg white powder samples with varying adulterant types and concentrations, this study constructed a dual model for the identification and quantitative prediction of adulterants using an improved one-dimensional convolutional neural network (1D-CNN). The qualitative model, which required no spectral preprocessing, exhibited accuracy rates (A_AR) of 98.19%, 99.38%, and 94.79% for bulking agents, nitrogen-rich compounds, and mixed adulterants, respectively. The overall A_AR reached 98.11%, with the lowest recognition concentrations (L_LRC) of 1%, 1%, and 5% for the three types of adulterants, respectively, and an average time spent (A_ATS) of 0.0177 s. For the quantitative model, detrending (DT) was used for spectral preprocessing to predict the concentration of bulking agents, while multiplicative scatter correction (MSC) was used for the concentration prediction of nitrogen-rich compounds and mixed adulterants. The determination coefficient of prediction (R_p²) of all three test sets exceeded 0.9, and the residual predictive deviation (RPD) was above 2.5, meeting the requirements of market regulation. The dual detection model provides key technical support for the development of portable near-infrared spectroscopy-based detectors.