Analysis of feature signals of electronic nose in honey nectar detection based on independent components analysis combined with genetic algorithm

Abstract: Focusing on the characteristics of electronic nose (e-nose), including fully, noisy and redundant, Independent Components Analysis (ICA) is proposed to extract differential signals of e-nose in the honey nectar detection (rape honey, linden honey and acacia honey). However， in order to match the principle of ICA (each vector standing for one observing signal) and overcome the shortage of ICA (the randomness of independent components), some transforms of ICA are needed to be carried out. Referring to the methods applied in brain image analysis, the research extends the signals of different samples in time direction. In this case, the order of independent components was assured. After the comparison of different number of ICA, which is evaluated by the accuracy of pattern recognition with the support vector machine model, the optimum number is confirmed as 8. Although the quantity data has been narrowed down dramatically, still 960 points are included, 120 points for each components. To the further simplification, Genetic Algorithm (GA) is used to select the characteristic points to remove the redundant information. 20 points, many of which are located in the absorption phase are selected. The results of GA selection show that although most of special points are located in the absorption part, there are still a part of points which are emerged in desorption part. In this case, only selecting the values in the peak is defective, which may lead to the ignorance of some special information which is included in other parts, like the desorption part. To testify the effective of the method, it is compared with other common processing methods, including raw data without being processed, maximum conductance of the original signals, the Principle Components Analysis (PCA), and data only processed by ICA without GA. The data are processed by these five different processing methods. After that, the Support Vector Machine (SVM) is employed as the pattern recognition method. Compared with other models, SVM model is built based on the principle of structural risk minimization inductive principle, which takes the empirical risk and practical risk into account at the same time. This consideration could be helpful for the small sample sizes detection, and can be benefit for the method to enlarge the area of application. The results show that method of ICA combined with GA obtained the highest predict accuracy 95.0% of which the accuracies of three different nectars were rape 24/25, linden 16/17, and acacia 36/38. Besides, the predict accuracy is close to the train accuracy of the model (96.3%) which means the model can provide better stability and better generalization. The study shows that the method can extract the differential signals and removed the redundant signals without affecting the ability of classification of the model.