A Novel Feature Selection Framework for Analyzing E-nose Data:Application to Evaluate the Quality of Chinese Dry-Cured Ham

Kang Qian,Chenyuan Wu,Jiaqing Zhang,Changqing Chen,Zhenbo Wei

DOI: https://doi.org/10.13031/aim.201900680

2019-01-01

Abstract:In this work, a portable home – made electronic nose was applied to differentiate and predict the grade of Chinese Dry – Cured Ham. To acquire more comprehensive olfactory information, six time – domain features and four frequency – domain features were extracted from every MOS sensor‘ response signals, but there are abundant irrelevant and redundancy information. Thus, this study proposed a hybrid filter – wrapper framework was used to select feature subset. The proposed method contains two steps: the Mutual Information Mixed Evaluation (MIME) worked as filter evaluation function based on Mutual Information, to eliminate those features which its MIME value is less than threshold value features rapidly; Support Vector Machine (SVM) classification was used as evaluation function, and the backward feature elimination with cross-validation (BFECV) algorithmic as an searching method, to remove redundancy features one by one during the iterative process. Then the hybrid filter-wrapper framework was compared with two feature selection methods. For grades distribution, the results indicated that hybrid filter-wrapper framework showed best performance. For qualitative prediction, Three classification algorithm (i.e., SVM , K-Nearest Neighbors and Logistic Regression classification algorithmic) were used to built prediction models based on feature subset which selected by MIME, SVM – BFECV and the proposed hybrid filter-wrapper method. The proposed method had the best mean prediction accuracy (95.13%) and lowest mean time consuming (16.35 s). Those results suggested that hybrid filter-wrapper framework outperforms the single feature selection algorithm, and could be efficiently used to analyse electronic nose data.