Identification method for different moldy degrees of maize using electronic nose coupled with multi-features fusion

Abstract: In this paper, in order to improve correct rate of discrimination result of different moldy degrees of maize using the electronic nose (E-nose), the influence of different feature combination representation types of E-nose signals on the discrimination result of moldy maize was studied in depth. In our investigation, the maize with 5 kinds of different moldy degrees was identification objects, and there were a total of 40 samples for each kind of moldy maize. Firstly, 30 samples were randomly selected from each kind of moldy maize for forming a training set (totaling 150 samples), and the rest 10 samples were used to form a corresponding test set (totaling 50 samples). To verify the robustness of this research finding, 5 groups of training sets and their corresponding test sets were randomly generated and respectively tested by the E-nose, and the test signals of the 5 groups of training sets and test sets were obtained; meanwhile their discrimination results were also compared with each other. Secondly, integral value (INV), average differential value (ADV) and relative steady-state average value (RSAV) of E-nose signals were extracted as 3 kinds of feature values; the five groups of training sets and corresponding test sets were respectively represented by each feature value, and also by their combination feature values. Then, the 5 groups of training sets were respectively analyzed by Fisher discriminant analysis (FDA) and 5 FDA analysis models were established, and then their corresponding test sets were used to verify the 5 FDA models. FDA results showed that: when the E-nose signals were represented by single feature or 2 features' combination, different moldy degrees of maize could not be discriminated effectively, but the correct rate of discrimination results based on 2 features' combination was better than that of the single feature, and the highest correct rate of single feature was 86%, while the highest correct rate of 2 features' combination was 96%; the identification ability of FDA was improved under the condition of 3 features' combination, the correct rate of discrimination result was at least up to 96%, and the highest correct rate of 3 features' combination was 100%. In addition, the feature representation difference of each sensor response signal was inspected with the help of Wilks ?-statistic, and the feature parameters of each sensor response signal based on 3 features' combination were selected and determined. FDA results displayed that the discrimination results of the maize with different moldy degrees before and after feature parameter selection were very similar, and the highest and the lowest correct rate based on feature parameter selection were 100% and 96%, respectively. So, it is necessary for the different sensors to be represented using different feature parameters so as to reflect their differences fully, and thereby the analysis complexity of E-nose can be reduced availably. The research finding clearly shows that the response signal of E-nose to moldy maize can be more effectively represented using multi-feature fusion, and the correct rate of discrimination result can be improved; at the same time, the research finding may not lose generality and provides a new idea of feature representation for E-nose signal.