GC–MS characterization of body odour for identification using artificial neural network classifiers fusion

• Human body odour characterization with GC–MS.
• Visual Discrimination of odour samples in feature space.
• ANN classification for odour recognition of individual.
• Fusion approaches for improved classification accuracy.

The focus of the present study is the human body odour recognition by analysis of information about the chemical compounds identified in their gas chromatography–mass spectrometry (GC–MS) chromatogram. The Artificial neural network (ANN) technique implemented in the current study, has been comprehensively used for classification and regression tasks in numerous applications. The experimental data set includes intensity characteristics (peak height, peak area, ratio of peak area and height) of several chemical compounds detected in GC–MS chromatogram of twenty odour samples (from four persons), and two non-body odour samples. The raw data set is transformed with logarithmic scaling, principal component analysis (PCA), and kernel principal component analysis (KPCA) in search of the better features by extracting. After preprocessing of data, feed forward back-propagation neural network (BPNN) technique is used in discrimination of body and non-body odour samples, as well as recognition of body odour to an individual. Although ANN classifier is optimized for the number of neurons, and training algorithms, the classification result is unstable and unsatisfactory (maximum correct classification rate 78% and minimum correct classification rate 44%). To improve the stability and accuracy of ANN classification results, data fusion approach is attempted. Eight different weighted and unweighted decision schemes of data fusion have been implemented in body odour recognition. Amongst them simple weighted vote (SWV), quadratic best worst weighted vote (QBWWV), and best worst weighted vote (BWWV) outperform with 100% class recognition outcomes, compared with a single ANN classifier.

Figure optionsDownload high-quality image (179 K)Download as PowerPoint slide