Qualitative and quantitative detection of camellia oil adulteration using electronic nose based on wavelet decomposition humidity correction
Dapeng Li, Han Jiang, Gan Yang, Zhongliang Gong, Tao Wen
Abstract
This study addresses the challenge of environmental humidity impacting the accuracy of metal oxide semiconductor (MOS)-based electronic nose systems in detecting adulterated camellia oil. A self-developed electronic nose platform with eight MOS sensors was used to detect adulteration in camellia oil mixed with varying ratios of rapeseed, soybean, and corn oils at different humidity levels (20%, 40%, and 60%). Wavelet decomposition using fourth-order Symlet wavelet was applied to correct response signals, mitigating humidity's effect on detection accuracy. Linear discriminant analysis (LDA), support vector machine (SVM), and random forest (RF) were employed to build qualitative identification models using pre- and post-correction datasets, while partial least squares regression (PLSR) and backpropagation neural network (BPNN) were used for quantitative prediction. Results showed that both qualitative and quantitative models significantly improved performance after correcting signal drift with wavelet decomposition. The qualitative model's 10-fold cross-validation prediction accuracy increased to 98.67% from 88.67%, while the quantitative model's average R 2 reached 0.99 with RMSE reduced to 3.64 mL/100 mL, compared to pre-correction R 2 of 0.86 and RMSE of 9.63 mL/100 mL. Notably, the RF and BPNN models outperformed others in qualitative and quantitative detection, respectively. These findings highlight the potential of electronic nose technology with humidity correction for authenticating camellia oil. • Developed an electronic nose platform for detecting camellia oil adulteration. • Used wavelet decomposition to correct humidity-induced signal drift. • Achieved 98.67% accuracy in qualitative detection after humidity correction. • Quantitative models showed R 2 of 0.99 and RMSE of 3.64 mL/100 mL post-correction. • RF and BPNN models outperformed others in qualitative and quantitative detection.