A method and system for analyzing continuous bit segments taken from a general data channel and classifying the sampled data by type, such as: voice, audio or data. B contiguous bits are converted to plus and minus deltas. The B-replaced values are then padded with B contiguous zeroes and the Fourier Transform of the padded sequence is computed. A power spectral density is derived from the Fourier Transform. In addition to the Fourier Transform, compression and entropy algorithms are performed on strings of bits within the message. The first B terms of the power spectral density and the results of the compression and entropy algorithms are used to differentiate and classify the data types, based on the premise that the combination of power spectral density, compression and entropy results yields parameters indicative of distinct types of data messages.

Method and system for classifying digital traffic

In this chapter we empirically compare neural networks and statistical methods on two telecommunications problems — one regression problem and one classification problem. Our goal is to explore the relative performance of neural networks and several statistical methodologies on moderate-sized regression and classification problems. A regression problem is one in which the goal is to estimate a numerical quantity based on partial input information; a classification problem is one in which the goal is to estimate the class to which an item belongs, again based on partial input information. We did not use a systematic method to choose problems for comparison, but rather chose one significant complex problem of each type. By significant, we mean that the underlying scientific or engineering question was important. By complex, we primarily mean that the dimensionality of the problem was large (roughly, greater than 20 measurement dimensions per observation or case).

Empirical Comparisons of Neural Networks and Statistical Methods for Classification and Regression

Current in-service nonintrusive measurement devices (INMDs) measure speech and noise levels, speech echo path loss, and speech echo path delay. Work on extending these measurements to include the automatic characterization of traffic on DS0-rate lines, i.e., 64 kbit/s PCM in the public switched network, is described. The ability to correctly identify speech and various types of data traffic could be used to gather more accurate statistics of local exchange traffic and could eventually serve to differentially route or tariff calls. In laboratory experiments, neural networks are used to distinguish speech from modem traffic, as well as differentiating between modems speeds. Classification is made on 14-ms segments of the signal with an overall accuracy of over 98%. The two-layered nonlinear neural networks performed better than both a linear-discrimination technique and a template-matching technique. >

http://ieeexplore.ieee.org/iel2/645/6841/00276606.pdf

C.M. Humphries

Papers

Characterizing DS0-rate traffic using neural networks