I found it there are functions
fann_scale_input( struct fann * ann, fann_type * input_vector)
fann_scale_output( struct fann * ann, fann_type * output_vector)
That scale input/output based on a previous trained data.
I guess that they should do the job.