New approaches to signal detection and classification

Date of Completion

January 2004


Engineering, Electronics and Electrical




Automated signal detection and classification has attracted considerable attention and has a broad application area. Of particular interest in this dissertation are long spectral line detection, information fusion of different waveforms in active sonar systems and automated processing (segmentation, re-estimation and classification) of time series. ^ In the first part of this dissertation (chapters 2 and 3), we deal with the long chirp detection and long smeared line detection. For the former case, the signal is long and weak, has an extremely slowly-decreasing frequency, and is corrupted by white Gaussian noise and possibly also by powerful tones. An online Hough transform detector is proposed. Its performance is analyzed, considering the detection, speed, and minimal detectable frequency slope. For chapter 3, a track-before-detect algorithm via the Hough transform and the probabilistic data association filter for a low-observable and long smeared line is presented. Small maneuvers are allowed for the line. This new algorithm works well at quite low signal to noise ratios. ^ In the second part (chapters 3 and 4), the fusion in sonar signal processing is addressed. Different waveforms can have complementary characteristics, it therefore makes sense to fuse the information from them. Four fusion schemes are investigated and the performance of fusion regarding the detection, estimation and estimation-tracking system is studied. It is concluded that fusion of the information of different waveforms can yield not only more robust performance, but in some cases outright preferable. Further, optimal detectors are derived for the fused “two-look” detection in clutter: the gain over traditional matched filter is enormous while the clutter to noise ratio is large. ^ In the third part (chapter 5), we focus on the automated processing of time series cast as variable duration hidden Markov modeled segments. A Gibbs sampling approach for segmentation is proposed; a joint segmentation and classification scheme is given based on a group of the Gibbs segmenters and the class-specific method; and an another Gibbs sampling approach is presented to re-estimate (i.e., train) the parameters of the time series. The class-specific features are employed for PDF estimation. ^