Assessment of the robustness of a neural network approach for the electromagnetic reconstruction of buried cylinders