Weighted Summation: Feature Extraction of Farm Pigsty Data for Electronic Nose

Abstract

When an electronic nose (e-nose) is used for prediction, extracting more useful information from the original response curve is of great importance. However, the most traditional feature extraction models in e-nose only sample a few data during the process of extracting features. To use more data and acquire more information to improve e-nose’s classification accuracy, we present a new feature extraction method called “weighted summation” (WS). In addition, this method was compared with other exiting methods, including maximum value of the steady-state response (MAX), curve fitting (CF), dynamic moments of the phase space (MD2), maximum value of the first-order derivative (Dmax), and Db1 wavelet transformation (WT). Dingfeng pig farm located at Changchun (Jilin Province, China) was used as odor source. Four kinds of odors taken from inside of pig barn in the morning and in the evening, and outside of pig barn in the morning and in the evening were used as the original response of e-nose. The reasons why we choose these four classes are as follows: to start with, the smell of the house has a great influence on the health of pigs; then, outdoor odors affect residents’ comfort level; and morning and evening are the most odorous hours. Experimental results demonstrated that for WS, MAX, CF, MD2, Dmax, and WT methods, accuracy in training set was 88.33%, 85%, 83.33%, 83.33%, 46.67%, and 51.67%, respectively, and accuracy in testing set was 100%, 100%, 91.67%, 91.67%, 41.67%, and 41.67%, respectively, suggesting that novel feature extraction method outperformed other methods. Moreover, a simple monitor system based on WS method was established to monitor the real environment in pig farm.