Signal classification using pattern matching of peak frequency magnitude histogram images
Roof, Aaron James
MetadataShow full item record
In this modern age of digital communications, new applications in machine to machine communication are being discovered every day, and the wireless spectrum is being flooded with a plethora of signals of varying format. The days of utilizing a single hardware device for only one specific signal are quickly being replaced by a "one size fits all" approach towards signal reception and processing. A Software Defined Radio (SDR) is just one such device in which a single piece of hardware is able to adapt its reception capabilities to match a variety of signal formats in order to receive a signal of any type. For such a device, the ability to process a selection of spectral space and identify a number of unknown signals is an important feature. In fact, for many SDR-like applications, signal classification is becoming an increasing necessary function. This dissertation introduces a new technique that can be used to identify key features required for signal classification. Many classification techniques assume that the received signal space is occupied by only one signal, and that the frequency of operation is known. However, in many systems, the receiver may be completely blind to the number and characteristics of signals within the bandwidth of interest. The method proposed herein does not assume there is only one signal in the space, nor does it require the frequency of operation of any signals to be known. The technique explored in this work proposes the collapsing of localized magnitude peaks from consecutive Discrete Fourier Transform (DFT) bins into histograms to create a two dimensional image of the frequency-magnitude density of the received signal space; referred to as a Peak Frequency Magnitude (PFM) Histogram image. This image can be a useful visualization tool in the characterization of the signal space in user assisted modes of classification, or the image could be used for machine classification by employing Automatic Signal Classification (ASC) techniques. One such ASC technique is proposed in this work which uses pattern matching of clusters in the image followed by a Decision Tree Classifier (DTC). A set of common signal types are generated for simulation stimulus in addition to using hardware to capture real world signals from a signal generator for validation of the technique. Two different forms of the basic algorithm are proposed and the tradeoffs compared. The limitations of the technique are also discussed as well as opportunities for future research to improve upon the performance of the algorithm.