To determine the relative status of the damaged floors of a building after an earthquake using building modal characteristics, instrumentation of the building is not required throughout its lifetime. In this paper, comparison of neural networks accuracy trained with fractional frequency change and mode shape change obtained from combination of three, four and five damage levels of different storeys of the building is evaluated. The network trained with combination of three damage levels of four and eight storey building is incapable of giving acceptable results. However, for four-storey building, network trained with four damage levels predicted good results and networks trained with five damage levels predicted excellent results. For eight-storied building, network trained with four damage levels gave acceptable results for storey level damage. The accuracy of damage severity, decrease with the increased number of building storey.