ART-2 is a self-organized and unsupervised artificial neural network constructed from adaptive resonance theory which can be used to classify continuous active data. But we found that the network has just used the big amplitude information during classifying the data by typical ART-2 network, especially the time series data. Therefore, we proposed an improved architecture based on typical ART-2. Then, we point out the superiority of improved ART-2 network over typical ART-2 network in theory. At last, a simulation is given to show the superiority