| Abstract: | The purpose of this paper is to present an automatic method of signal analysis. To help physicians in their diagnostics, this method is implemented on a minicomputer in order to detect non-stationary points in electroencephalograms. The signal is modelled with an autoregressive filter. The parameters of this filter are adapted at each step. Identification gives the best model in the sense of a cost function representing the mean square error of noise, which is estimated during the optimisation time-window. The cost function is expressed by a quadratic formula. This allows the use of a fast algorithm, the 'conjugate gradient method'. An original statistical test is developed to detect non-stationary points in the signal. The performance of this method is tested with artificial data to determine the sensitivity of method parameters. Detection using real data is presented. |
