Research On Online Drift Counteraction Of Electronic Noses In An Active Learning Framework

In recent years,electronic nose(E-nose)technology has gradually developed steadily and matured.E-nose is a typical artificial olfactory system which is good at identifying volatile compounds in odor.And E-nose is mainly composed of a gas sensor array,a signal pre-processing unit and a pattern recognition unit.Among them,the sensor array is the basis of the E-nose system.However,online drift of the sensor response is common in practical applications.The sensor drift will cause the compatibility between the response and the algorithm model to deteriorate over time,which will lead to the degradation of E-nose performance.At present,there are many studies using different methods to solve the inevitable drift problem in the development of E-nose system technology.However,in the online drift-suppression scenario,the pre-label correction method is difficult to meet the requirement of obtaining a large number of samples with comprehensive categories.Simultaneously,due to the accuracies of the labels given by the model are less in the weak label calibration method,it is impossible to generate an excellent correction set.In addition,the correction method without lables obtains drift information from unlabeled drift samples.However,the model performance cannot be significantly improved due to the low information richness.In order to obtain a high-accuracy and high-quality correction set with as little human intervention as possible,this study introduced an Active Learning(AL)method to complete the online update of the recognition model.In the online drift-suppression scenario of the E-nose system,in order to ensure the improvement of model performance,the drift correction set needs to meet the requirements which are comprehensiveness of categories and high representativeness of sample distribution.Therefore,this study contains three algorithms based on active learning framework.Specifically,the work and results obtained in this thesis mainly include:(1)E-nose hardware platform implementation and data collectionThis research designed an experimental platform for E-nose data acquisition,which mainly includes a control unit,a sensing unit and a gas sampling unit.At the same time,the online drift data set was collected using the E-nose system.(2)An active learning on adaptive confidence rule has been proposed and verifiedAiming at the problem that traditional active learning sample selection strategies cannot quickly cover the drift sample distribution,this thesis describes a method named Active Learning on Adaptive Confidence Rule(AL-ACR)based on the multi-sample selection criteria.The core idea is to adaptively adjust the instance selection criteria in order to the selected instance can cover the drift data as widely as possible.(3)An active learning on dynamic clustering sampling has been proposed and verifiedAiming at the problem of incomplete categories of drift correction datasets obtained by previous active learning methods,this study considers the sample category information while calculating the value of sample information.Active Learning on Dynamic Clustering Sampling(AL-DCS)method has been proposed in this thesis.The method is divided into three stages: initialization,clustering and instance selection.(4)An active learning on classifier-state sampling has been proposed and verifiedIn order to avoid the performance of the active learning method being affected by the distribution of the drift samples themselves,this thesis describes a method named Active Learning on Classifier-State Sampling(AL-CSS)based on classifier state sampling.At the same time,in order to reduce the running time of the algorithm,AL-CSS based on training-sample refining(AL-CSS-R)method has been proposed.In this study,two data sets are used to verify the effectiveness and practicability of the three methods respectively.From the perspective of algorithm recognition accuracy and algorithm recognition efficiency,they show that three methods perform superior anti-drift ability in the long and short term.Meanwhile,the above experiment illustrates that the active learning method is suitable for the online drift-suppression scenario,and can effectively solve the problem of the sensor’s online drift in the E-nose.Among them,The AL-ACR and AL-DCS methods are suitable for scenarios with high requirements on runtime and accuracy.The AL-CSS and AL-CSS-R methods are suitable for scenarios that require high algorithm accuracy but low runtime.