Clustering is a particular subset of the data analysis methods, which aims at researching and discriminating groups (clusters) of similar observations in a dataset. Often these data can be observed at different step forming time series as the evolution of a stock market value or meteorological phenomena. In some time series, the sequences of observations exhibit distinct and interpretable phases, which we call "regimes". For instance, a car speed can show acceleration, cruise speed and breaking phases. In this article we propose a method dedicated to the clustering of this particular kind of time series. It consists in the combination of three steps: an individual segmentation of the time series, the construction of a common regimes dictionary, and the final clustering of categorical sequences produced from the recoding of the time series in this dictionary. We present the different advantages of this method and the results obtained on several public datasets.