Adaptive Bayesian spectrum estimation

A method is described for the numerical analysis of nonstationary signals whose spectral distributions vary slowly enough with time for an adaptive method to be used. The data is processed by blocks. The theory describes the signal by an autoregressive series of high order with stochastic coefficients having a Gaussian distribution. The coefficients are considered as state variables and can thus be estimated in real time by a special fast Kalman filter. This raises the problem of determining an essential tuning parameter which is a regularization coefficient. This is done according to maximum likelihood. Numerical experiments on synthetic signals have demonstrated the feasibility of the method.<
>