Improvement of recognition capability of fuzzy-neuro LVQ using fuzzy eigen decomposition for discriminating three-mixture fragrances odor

Artificial odor recognition system is developed for automation of detection and classifications of aromas, fragrances, vapors and gases. We have developed various artificial neural networks algorithms used as the pattern classifier for recognizing mixture fragrances, including the family of fuzzy-neuro LVQ (FNL VQ) algorithms. The developed neural networks classifiers however, show low recognition rate when it was used to recognize three-mixture fragrances problems. There are still major difficulties in the usage of FNLVQ algorithms, i.e., choosing the initialization of the fuzzy-reference vectors. The initialization step is important due to different selections of the initial reference vectors may potentially lead to different partition for different classes, which hampered the superiority of the algorithm. In present study, we proposed a novel initialization method, i.e., by transforming all the fuzzy vectors from the original aroma space into its eigenspace prior the usage of FNLVQ. Experiments are conducted using our odor recognition system and the performance of FNLVQ in eigenspace shows higher recognition rate compare with that in the aroma space, especially for 18 classes of three-mixture fragrances odor problem.