Adhesive-emitted odorants detection using an electronic nose: Unraveling algorithm applicability with controlled dataset characteristics

Abstract

Electronic nose technology is becoming increasingly important in pollution monitoring; however, the inappropriate selection of pattern recognition algorithms may lead to performance decreases. In this study, an electronic nose was developed for the qualitative classification of adhesives and quantitative detection of adhesive-emitted odorant concentrations. Meanwhile, the applicability of commonly used pattern recognition algorithms (support vector regression, partial least squares regression, artificial neural network (ANN), random forest regression (RFR), ridge regression, and Lasso regression) to datasets with controlled volumes and interference intensities was investigated by comparing their quantitative performance. In qualitative analysis, the support vector machine with polynomial nonlinear kernel and random forest achieved 100% accurate classification of 11 adhesive samples. In quantitative analysis, RFR demonstrated good generalization ability and interference insensitivity, with a mean absolute percentage error (MAPE) below 3% in the large-volume strongly interfering dataset and less than 25% in the small-volume strongly interfering dataset. While ANN showed certain data volume dependence and interference sensitivity. For dataset with small volume and weak interference, algorithms with simple structures could achieve accurate quantification (MAPE of 10%). Moreover, the algorithm applicability was validated on homologous datasets and the effectiveness of cluster analysis to remove outliers was discussed.