Robust Spectral Classification
Spectral classification is a commonly used technique for discriminating between two or more signals. The first step in the classification process is to sample a signal with an analog-to-digital converter. Then the power spectral density is estimated. To classify the data, the estimated power spectral density of the unknown signal is compared to power spectral densities from two or more known templates. The comparison is done using a classifier. Despite the substantial prior research effort put into developing a robust classifier, the results are not great and in some instances are not even satisfactory. ^ The topic of this thesis is to evaluate a classifier that may be more robust than those currently used; the realizable Poisson likelihood function. Robustness is determined by the probability of correct classification when there are differences between training data and observed data. Taking the familiar form of the Kullback-Leiber divergence, the realizable Poisson likelihood function is mathematically tractable since it is derived from an alternative model for the power spectral density of a non-homogeneous Poisson process. ^ The realizable Poisson likelihood function was compared to other popular classifiers. Monte Carlo simulations were done using autoregressive processes with and without distortions added to the observed data. Then a more thorough analysis was done using actual data. Results are presented that show the realizable Poisson likelihood function to be a robust classifier. The performance of the realizable Poisson likelihood function deceases only very slightly with moderate signal-to-noise ratios and in the presence of channel distortions. This is compared to significant performance decreases of other classifiers.^
"Robust Spectral Classification"
Dissertations and Master's Theses (Campus Access).