QBC-Softmax Algorithm for E-nose Data Processing Based on Different Informativeness Evaluations

An electronic nose is an intelligent system for recognizing various gases, which consists of a gas sensor array and patter recognition algorithms. In general, getting a well performing model requires a large number of labeled samples. However, labeled samples are always difficult, time-consuming or expensive to obtain. In fact, it is far easier to collect unlabeled samples than labeled ones. In this paper, one of the active learning algorithms, an improved query by committee algorithm for the softmax classifier, is applied to E-nose to reduce the size of samples that needs to be labeled. Firstly, three informativeness evaluations (Jensen-Shannon divergence, Hellinger divergence and Kellback-Leibler divergence) are applied to compute the divergence degree among the committee members' vote results for query samples. Secondly, a pool-based sampling strategy is adopted to pick out the informative query samples, which considers the average divergence degree of the whole query samples as the threshold. The results have shown that the proposed method performed better than the random selection method. Specially, when the size of the initial training set is small, Hellinger divergence has the best performance. The Jensen Shannon divergence and Kellback-Leibler divergence perform better with the increasing of size of the train set.