Study On Class Balancing Method For Online Drift Correction Of An Electronic Nose By Active Learning

Artificial olfactory system(electronic nose)is an odor-detection device based on artificial intelligence algorithm.In the popularization process of the device,it is particularly important to solve the drift problem under the on-line continuous conditions.Considering the feasibility and characteristics of calibration samples,active learning is chosen as the basic framework of online drift calibration method in an electronic nose.In this framework,due to the randomness of labels caused by the fact that drift calibration samples are taken from previous online samples,it is likely to appear the "class imbalance" problem.In this study,two contents of research have been carried out on this issue.Firstly,in order to improve the drift correction effect of minor-category samples in the active learning,an improved active learning sample selection method,(Query By Committee for Balancing Classes,QBCBC),has been proposed in this part.QBCBC focuses on selecting and labeling samples that critically impact the classification of minor-category samples,as well as improves the recognition ability of trained classifiers for each sample category.Based on the "query by committee" strategy,QBCBC constructs two committee members using integrated classifiers on bootstrap-sampling by the “maximum voting divergence” criterion.In view of the class-imbalance problem possibly happened in calibration datasets,selecting the samples to label by the sub-classifier outputs of committee members related to the scarce class with the greatest divergence.The aim is to select as many samples as possible at the classification boundary of the scarce class and make the class information of the online calibration samples balanced.The experimental results show that QBCBC can comprehensively improve the recognition accuracy of the classifier on all kinds of samples.In addition,when the referenced class imbalance ratio is set between 1.2 and 2,the proposed method can play a significant role in improving the overall accuracy(G-mean)up to 7%.Secondly,to solve the contradiction on limited storage resources and unlimited increase of correction samples under active learning framework,an Inducting-Rebuilding Function Fitting(IRFF)method has been proposed for calibration samples based on function fitting.IRFF obtains the labeled calibration samples online in active learning framework and uses Gauss function to fit and summarize calibration-sample information.When calibration is needed,various calibration samples can be reconstructed from the fitted Gauss function.For the number of the reconstructed calibration samples is not limited by the amount of real samples,the usage of IRFF can avoid class imbalance in calibration samples.In addition,the information of samples is transformed into Gauss distribution parameters(mean and variance)in the process of “induction”,which achieves the effect of compressing the storage space of calibration samples.The experimental results show that IRFF reduces the storage space of calibration samples at the cost of a certain decline of recognition rates.On the electronic nose drift dataset used in this study,when the number of reconstructed samples is kept between 10 and 60,the compression ratio of storage space for samples can reach 27:1,and the average G-mean can be maintained at about 0.75.In this dissertation,the theory and method of active learning are introduced into the work of electronic nose drift suppression.Aiming at the problem of unbalanced calibration samples and limited storage space,the methods based on sample selection strategy and function fitting are proposed.The works of this dissertation lay a foundation for further research on online drift suppression methods of artificial olfactory system.