We consider a specific case of the anomaly detection problem where the considered system is unknown but observable. The aim is to detect if this system changes or not for a long period, and, if possible, to characterize these evolutions. The proposed method consists of two steps. The first one is to learn the system thanks to a Hidden Markov Model with explicit duration of states, using observations of the system at different moments. The second step is to compute distances between these models to detect the evolutions. Due to the constraints of the learning algorithm and computational time, we proposed a new method to compute distances between Markov models, based on the Kullback-Leibler divergences. The method is tested on a trial evolving model. These results are .