In the present work, a digital twin of a radial-axial ring rolling machine was built by modelling the time series of the positions of the tools and control signals rather than the metrics of the produced rings, as performed in previous studies. Real data from the industry was used for modelling. The used model selection methodology is shown in detail to replicate such work for similar systems in the steel industry. The modelling results of ARX, ARMAX and orthonormal basis model structures are shown; additionally, they were validated considering SISO and MIMO systems. The modelling results were better when the subsystems considered were ARMAX and MISO than when ARX and SISO were taken into consideration. The best modelling results were obtained when physical knowledge was included in the model structure. Lastly, it was found that the model error of the horizontal subsystem could be used for predictive maintenance.