Optimized KPCA method for chemical vapor class recognition by SAW sensor array response analysis

This paper confirms the suitability of kernel principal component analysis (KPCA) as a robust feature extraction and denoising method in sensor array based vapor detection system (E-nose). Particularly the study focuses on response analysis of surface acoustic wave (SAW) sensor array in chemical class recognition of volatile organic compounds (VOCs). Usually KPCA results deprived performance compare to linear feature extraction methods. However its performance is affected by the proper selection of kernel function and optimization of allied parameters. We have presented the comparative performance analysis of feature vectors extracted by KPCA method using five types of kernel functions in combination with support vector machine (SVM) classifier. Study outcomes are based on analysis of 12 data sets (enclosing different intensity of additive noise and outliers) generated with SAW sensor model simulator. We find that in research of kernel function selection; polynomial kernel achieves persistently maximum class recognition rate of VOCs (average 82 %) even in presence of high level of additive Gaussian noise and outlier and anova kernel results minimum class recognition rate (average 70 %). The class recognition efficiency of feature vectors extracted by rest of the three kernel functions lies in between these two.