Intelligent techniques for spectral estimation

The application of artificial intelligence (Al) techniques to improve spectral estimation techniques is examined in the restricted environment of biomedical analysis. In many cases the fast Fourier transform (FFT) periodogram is applied because it is simple to use and it provides an unambiguous output. There are a variety of parametric spectral analysis algorithms which can improve upon the periodogram results, but they usually require certain critical initial parameters to be accurately specified. This is typically done by trial and error with the user successively analysing his results and repeating the process with revised parameters. The use of a Prolog rule based expert system is investigated to determine the correct autoregressive (AR) model order. This is achieved by searching for a region of stability in the outputs of AR estimators and through comparison with the FFT based periodogram. This provides spurious peak compensation capabilities and subsequently permits the peaks to be located and assigned a confidence level by the artificial intelligence (Al) system. Our application is restricted initially to estimating the fundamental content of a fetal heart signal, which is a harmonic process where the signal is stationary only over short (32 sample) data records. We further report an Al algorithm to estimate and track the fundamental frequency of real fetal signals over 32 and 64 point data records with AR filter orders up to 30 which permits the analysis of up to 1 5 simultaneous sinusoids.