Adaptive spectra estimation of non-stationary underwater sound reverberation process

The spectra of acoustic Dopper current profiler (ADCP) signals which are a non-stationary underwater sound reverberation process were estimated by an adaptive lattice filter. Three programs, i.e., stochastic gradient (SG), fast Kalman (FK) and recursive least square (RLS), were used. Learning curves were given in the case of optimum parameters. From these curves it can be concluded that, RLS program is the best one. Its convergence speed is comparable with that of FK, but its mean square error (MSE) is smaller than that of FK and SG. The spectra of ADCP signal were estimated with RLS program. The number of power spectrum peaks may be one, three or two with increasing nonstationarity of medium movement, which correspond to the quasi-stationary area, non-stationary area and strong non-stationary area of medium movement, respectively. They are consistent with the theory of evolutionary spectra and Wigner-Ville spectra of moving medium. In these theories, both the non-stationary factor A and characteristic width B are significant. When the time interval T corresponding to the volume element of the scatters is smaller than B, it can be supposed that the power spectrum displays a single peak characteristic, when T is in the neighbour of B, there are three peaks in power spectrum, and when T is larger than B, the power spectrum which displays a two peak characteristics can be adopted.