AbstractIn this paper we develop algorithms, resting on the principle of Recursive Least-Squares, for the treatment of multichannel signals. A method of modelling in packs is presented in the scalar and vectorial cases, the latter of which takes into account the independent or dependent context of the channels. In a situation of slowly evolving signals, a detection method of model changes with two thresholds is proposed. A simulation of the presented algorithms is done on electroencephalographic signals on which our method was successfully applied to the detection of typical epileptic events.