Adaptive filtering and the identification of tones in broadband noise

This thesis addresses the sonar problem of identifying tones contaminated by additive noise and investigates the use of adaptive filtering techniques as a solution. The work carried out falls within a framework that involves time-frequency distributions and parametric estimation methods. The original finite impulse response - least-mean-squares (LMS) - adaptive line enhancer (ALE) is examined. Attention is focussed on a drawback in on-line performance: the appearance of spurious, misleading streaks in the braodband regions of the associated lofargram display. An analysis of the time variation of the LMS ALE's frequency response is conducted, and the streaks are attributed to the ALE's slowly changing frequency response. Infinite impulse response (IIR) adaptive line enhancement has been considered in terms of the second-order ALE. Five independently produced IIR ALE schemes have been studied: the filter structures and algorithms used and tonal estimation are considered for moderate and low signal-to-noise ratios (SNRs). It is observed that for low SNRs the geometry of the associated performance function adversely affects search performance. The performance of a particular IIR ALE has been improved by the introduction of a `state-dependent' pole radius. An adaptive scheme for generating the Gabor time-frequency distribution has also been formulated and studied. A method for analysing this scheme has been suggested whereby the adaptive mechanism is expressed in terms of transfer functions. Although the analysis method is restricted to small-sized (Gabor) distributions, the adaptive scheme itself is applicable for larger distributions. This has been demonstrated by means of simulations. In addition a brief comparison with the Short-time Fourier transform has been made with regard to the depiction of noise-corrupted tones.