A digital twin of a mechanical system (a pair of axial rolls in a ring mill used in a steel plant) with poles close to the unit circle and the real axis in the discrete pole-zero map was built. The system was excited by a signal concentrated in the low-frequency band. For this particular case, it is shown that the ad-hoc combination of ARMAX and orthonormal basis filter model structures outperform model structures based on either ARMAX or orthonormal basis functions when estimating the poles of the basis by analyzing the data in the frequency domain. The followed modelling methodology of the system is described in detail to help replicate the work for similar systems in the steel industry. Real production data from a steel plant were used in contrast to previous studies, where the combination of ARX and ARMAX with orthonormal basis filter model structures was evaluated using simulated data instead of real data. We believe that the resultant model can be used when having systems with poles close to the unit circle and real axis and poor excited input signal concentrated in the low frequency band. The resultant model can be used for condition monitoring and failure detection.