Rapid Detection and Prediction Models for Harmful Substances in Food Based on Support Vector Machine

Facing the challenges of diverse types and complex matrices of harmful substances in food, and the time-consuming nature of traditional detection methods, which are insufficient for rapid on-site screening, this study proposes a rapid detection and prediction model based on Support Vector Machine (SVM). A standardized dataset covering both liquid and solid foods is constructed. Raw spectral data is collected using a near-infrared spectrometer, and the concentration of harmful substances is determined using national standard methods as the gold standard label. Secondly, spectral preprocessing is performed through standard normal variable transformation and Savitzky-Golay first derivative, and wavelet transform and principal component analysis are combined to extract key wavelength features strongly correlated with the concentration of harmful substances. Based on this, an SVM classification and regression model with radial basis functions as the kernel is constructed, and parameters are optimized through grid search combined with five-fold cross-validation. Experimental results show that the proposed model achieves an average detection accuracy of 96.1% on nine types of food matrices, with an average detection time of only 2.9 minutes. The regression prediction coefficients are generally higher than 0.98, and the root mean square error is lower than 0.02 mg/kg, providing a feasible solution for rapid on-site screening and risk warning of food safety.