Classification of transmitter transients using fractal measures and probabilistic neural networks
Shaw, Donald B.
This thesis presents a method of identifying the source of radio transmissions by analysis of the transients exhibited at the start of the transmitted signal. It is motivated by the intriguing possibility of identifying radio transmitters used in violation of federal and international regulations. As well, such a system could be directly used for analysis or classification of other nonstationary signals such as speech or power system transients. The system developed in this thesis uses a multifractal analysis for precise segmentation of a transmitter transient from the ambient channel noise. This is critical to ensure that the portion of the signal being analysed does not contain meaningless noise and, at the same time, represents the entire transition from noise to signal. Then, using a similar multifractal method, significant features of the transient are extracted and stored for neural network analysis. This modelling process is equally important as it provides a means to reduce the size of the data for efficient neural network processing, while providing significant emphasis on the most important features. Finally, the transient model is classified using a Probabilistic Neural Network (PNN). Experimental results indicate that this classification system is fast and accurate. The three stages of segmentation, feature extraction, and classification are performed in about a half second for a 16 kB transient. In the most successful experiment, the system was trained with 160 out of the 415 available transients, representing eight different classes of transmitters. Testing the system with the remaining 255 transients yielded results in which 96.9% of them were classified correctly.