Study On The Multiple Classifiers Ensemble Technology In Air Quality Monitoring System

Electronic nose (E-nose) is an application of the artificial olfaction analogtechnology. By imitating biological olfactory, it can percept, analyse and identifyobjects to be tested. First, it gets information through a sensor array and then usespattern recognition analysis to detect gas type and predict concentration. The E-nose hasthe characteristics of rapid detection and real-time online lossless or non-invasivemonitoring. Taking the application of E-nose in air quality monitoring as the researchbackground, this thesis focuses on the study of the software design, feature extractionalgorithm and pattern recognition algorithm.The air quality monitoring E-nose system mainly contains hardware system,software system and pattern recognition algorithm. It can detect six kinds of gases:formaldehyde, benzene, toluene, carbon monoxide, nitrogen dioxide and ammonia bothqualitatively and quantitatively.At first, the author aims at the research of feature extraction which is involved inthe field of pattern recognition. The target contaminants’ pattern is linearly inseparablein E-nose system, and then it is difficult to extract the nonlinear features with generallinear feature extraction method such as: Principal Component Analysis (PCA),Independent Component Analysis (ICA). In order to sovle this problem, this thesisfocuses on nonlinear feature extraction method: Kernel Principal Component Analysis(KPCA) and proposes the application of KPCA to Support Vector Machine (SVM) forfeature extraction. Also, it compares favorably with PCA and ICA.In the meantime, a lot of works has been made to improve the classificationaccuracy of contaminate gas. A novel ensemble classifier model which solves amulti-class recognition problem in E-nose is proposed in this paper and it aims toimprove the accuracy and robustness of classification. First, KPCA method is used fornonlinear feature extraction of E-nose data; second, E-nose data after feature extractionare used as inputs of SVM to estabilish five base classifiers; finally, in the process ofestablishing classifiers ensemble, an effective fusion strategy was presented to integratethe decisions from the base classifiers, and the fusion result is the final output of theE-nose system. Two methods of fusion strategy were studied: majority voting (MV) andan improved weight approach in which weights assignment is done based on thepredictive accuracy of each base classifier on each gas. In addition, the performance of the proposed classification model was compared with that of base classifiers onvalidation dataset. Simulated with the E-nose data, experimental results show that,average recognition accuracy has achieved92%which are higher than86%obtainedusing base classifiers. The results of this work demonstrate that the proposed model canachieve a better performance both in recognition accuracy and generation comparedwith that of base classifiers.