Parameter Optimization of Support Vector Machine to Improve the Predictive Performance for Determination of Aflatoxin B(1) in Peanuts by Olfactory Visualization Technique

This study proposes a novel method for detection of aflatoxin B(1) (AFB(1)) in peanuts using olfactory visualization technique. First, 12 kinds of chemical dyes were selected to prepare a colorimetric sensor to assemble olfactory visualization system, which was used to collect the odor characteristic information of peanut samples. Then, genetic algorithm (GA) with back propagation neural network (BPNN) as the regressor was used to optimize the color component of the preprocessed sensor feature image. Support vector regression (SVR) quantitative analysis model was constructed by using the optimized combination of characteristic color components to achieve determination of the AFB(1) in peanuts. In this process, the optimization performance of grid search (GS) algorithm and sparrow search algorithm (SSA) on SVR parameter was compared. Compared with GS-SVR model, the model performance of SSA-SVR was better. The results showed that the SSA-SVR model with the combination of seven characteristic color components obtained the best prediction effect. Its correlation coefficients of prediction (R(P)) reached 0.91. The root mean square error of prediction (RMSEP) was 5.7 μg·kg(−1), and ratio performance deviation (RPD) value was 2.4. The results indicate that it is reliable to use the colorimetric sensor array with strong specificity for the determination of the AFB(1) in peanuts. In addition, it is necessary to properly optimize the parameters of the prediction model, which can obviously improve the generalization performance of the multivariable model.