This study was motivated by the characterization of the dynamic evolution of some critical railway components (point machines and door systems) using condition measurements acquired through embedded sensors. Its final objective is to build a decision-aided support for their preventive maintenance. One of the difficulties in achieving this goal is that, during their dynamic evolution, these components may switch between different states due to various operating contexts (different hygrometric conditions, different levels of train inclinations). We propose to solve this problem by automatically extracting, from temporal data, clusters whose characteristics evolve over time. In this framework, the clusters can be interpreted as the states within the operating contexts. This dynamical clustering problem is addressed by assuming that the data are distributed according to a mixture of Gaussian distributions whose centres’ are themselves distributed according to Gaussian random walks. The resulting model can be seen as a mixture of state-space models. The parameters of the proposed model are estimated by maximizing the likelihood function via the Expectation-Maximization algorithm. The preliminary results on both simulated and real data show the ability of the proposed model to accurately estimate the parameters while keeping a low clustering error rate.