On the Use of Numerical Models for Training Supervised Learning Algorithms for SHM