State estimation and dynamical model identification from the observed data have attracted much research effort during recent years. In this paper, an identification method of a system based on the unscented Kalman filter (UKF) and group method of data handing (GMDH)-type neural network is introduced and applied. Probabilistic metrics, instead of deterministic metrics, are used to obtain a robust Pareto multi-objective optimum design of the UKF-based GMDH-type neural network. The simulation results show that the UKF-based training algorithm performs well in modelling some explosive cutting and forming processes, and exhibited more robustness in comparison with those using a traditional GMDH-type neural network.