I'm not sure about your specific application or if this answers your question, but you may want to consider multiplying your FFT output by the sample period (1/sample rate).  This will give you a more consistent scaling for your FFT output for different sample rates.  Otherwise, consider using a unitary DFT, which is usually just an additional scaling of 1/sqrt(N) that you apply to your FFT output (or input), where N is the length of the input signal in samples.  You may also have to consider other rescalings if zero-padding is applied..

Thanks. I am working on Acoustic Emission signals. They are characterised by low amplitudes and High frequencies. good thing is I used  a consistent sampling rate of 2millionj samples per second.

May be I do not understand well your question, but this could be typical problem of frequency analysis. First of all the Parselval's theorem must be valid, i.e. the energy (or power) obtained by integrating the squared signal in time domain must be the same as sum of squared amplitudes of your spectrum in frequency domain. This fact results in effect that doubling the frequency resolution reduces the levels of spectral lines of 3 dB.

I think the key issue is whether your spectral subtraction algorithm preserves phase. The simple ones just operate on magnitude spectra. I am not aware whether there are any algorithms accurately preserving phase. If this condition is not met, the coincidences of erronous phases can easily produce local increases in the amplitude of your output signal (after the ifft). However, whatever the phases are, the RMS of the output signal should not exceed that of the input signal, on the ground of the Parseval's theorem, referred to above by Ondrej Jiricek. If this is not the case, you are likely to have an error somewhere in your procedure.

Thanks. The answers were helpful