This work deals with the dynamic behaviour of a lean premixed gas turbine combustor. The study aims to achieve a classification of experimental burner dynamic behaviour and is based on the geometrical properties of the attractors of the system variables. Several experiments were performed varying the flame stoichiometric ratio land the pilot fuel percentage PFP. The dynamics of the experimental time series of the flame front heat release were described by using vectors collecting information on the topological distribution of the attractors. Therefore. unsupervised Kohonen associative memories were trained to create clusters of operating conditions characterised by similar dynamical behaviours. Kohonen associative memories were able to divide the experimental operating conditions into different clusters according to the different values of the flame stoichiometric ratio. The results of the clustering underline the possibility of being able to define an algorithm for combustion-instability pattern recognition that takes into account the highly non-linear effects which govern combustion processes.