In this paper we considered a streaming data classification problem. First we introduced a concept of upper and lower envelopes of time series in order to reduce dimensionality of them. Next we merged machine learning tools like feedforward neural networks for selection principal attributes as well as decision rules of the form if … then … for time series classification. In result a novel representation of time series was obtained characterized by high dimension reduction and no false classification. A numerical example is performed showing 100 % accuracy.