Multivariate autoregressive feature extraction and the recognition of multichannel waveforms

It is proposed that the autoregressive coefficient matrices appearing in a multivariate autoregressive model fitting may be used for feature extraction purposes in problems concerning recognition of multichannel waveforms. It is demonstrated how the information contained in the autoregressive parameters may be further compressed by applying the ordinary or a modified Karhunen-Loeve expansion. The feature extraction procedures are illustrated on a large data base of seismic wave traces originating from shallow earthquakes and underground nuclear explosions. The results obtained (using a multivariate Gaussian classification algorithm) suggest that the combined autore-gressive/Karhunen-Loeve method has a considerably larger discrimination potential than the more conventional seismic discriminants.