A lot of researches on classifiers, which can perform well with a given set of feature vectors, have been done. However, researches on feature vectors, which extract better feature vectors automatically, have not been done very much. We face two problems when we consider feature extraction process. One is how we can make a good feature extractor, and the other is what more separable features are. In this paper, we solved these two problems by proposing feature extractor-training methodology that uses new cluster validity as an objective function. By combining feature extractor to Fuzzy Neural Network Model, we achieve on-line adaptation capability as well as optimized feature extraction. The result shows recognition rate of 97% when on-line adaptation is being done.