Modelling of non stationary systems based on a dynamical decision space

Overview & Synopsis

A new approach based on pattern recognition techniques and dedicated to the monitoring of non stationary systems is presented in this paper. More precisely, it consists of a recursive subspace identification algorithm combined with an adaptive classifier set for non stationary environment. The system identification method which provides a recursive estimation of a linear state space model is firstly described. Then, a feature vector representing the system functioning state is extracted from this estimated model. Next, the dynamical clustering algorithm which online learns the functioning modes and continuously determines the current mode of the system is introduced. Its auto adaptive and unsupervised abilities to take into account system modes evolutions are finally emphasized on simulation examples.