Quantitative detection of talcum powder in wheat flour based on near-infrared spectroscopy and hybrid feature selection

Highlights • Near-infrared spectroscopy technology was used to detect the content of talcum powder in wheat flour. • LOF method was used to detect abnormal samples, and SPXY method was used to divide the sample set. • The GBDT model, Adaboost model and LightGBM model were established to predict talcum powder content in wheat flour. • The hybrid feature method combining EN and GA was used to select effective feature. Excessive talcum powder in wheat flour would bring great harm to the health of consumers. How to quickly and accurately detect the content of talcum powder in wheat flour was of great significance. In this study, based on the advantages of near-infrared spectroscopy (NIRS) technology in material detection, the talcum powder in wheat flour sample was quantitatively detected. In this study, 123 wheat flour samples mixed with different content of talcum powder were prepared based on three types of wheat flour, and 41 samples were configured for each type of wheat flour. Among them, the content of talcum powder in 41 samples configured from each type of wheat flour ranged from 0% to 20%, and the content gradient was 0.5%. Firstly, local outlier factor (LOF) was used to eliminate two abnormal samples, and the remaining samples were divided into 85 training set samples and 36 prediction set samples according to the sample set partitioning based on joint X-Y distances (SPXY). Then, the performance of various spectral preprocessing methods and their combinations were compared. Among them, the performance of gradient boosted decision tree (GBDT) combined with standard normal transform (SNV) and first derivative (1D) was proved to be the best. Then, the effective features were selected according to elastic net (EN), genetic algorithm (GA) and EN + GA. Among them, EN + GA was proved to be the best, and 55 effective features were selected from 1050 features. Finally, the detection model was established to predict the talcum powder content in wheat flour. Through the evaluation of external samples, the correlation coefficient (R 2 ), root mean square error of prediction (RMSEP) and relative percent difference (RPD) of the detection model on 20 new samples with low-content talcum powder reached 0.9242, 1.3185 and 3.3443 respectively. The results showed that this study provided a new idea for the efficient, nondestructive and rapid detection of talcum powder in wheat flour, and had adaptability and practicability for low-content talcum powder samples. At the same time, the hybrid feature selection method used in this study was effective and feasible.