Progress in the field of Structural Health Monitoring (SHM) includes applications of model data approaches with numerically generated responses originating from Finite Element (FE) simulations. The simulated data may be used for supervised training of Machine Learning classifiers to perform damage identification on some later experimental state. The reliability of the numerically trained classifiers depends on the quality of the training data in terms of the simulation model error contained. Even though FE models can be updated on the intact state for better accordance with the experiment, some model error still remains for most complex structures. This error can limit the ability of generalization to the unmeasured yet damaged experimental states depending on how distinct they are. In the present work, the effect of model error on subsequent classification of experimental damaged states is tested for a lab-scale bridge truss. A Convolutional Neural Network classifier is applied, trained by FE responses. Results show that larger and more distinct damages can be classified with more accuracy compared to small and less distinct, with the latter showing higher prediction bias. The numerical data reliability is found to be reflected on the intact state numerical and experimental feature map shifts.