I have seen in a few papers, for instance in this one, the following method :
- Find an initial "frequency of interest" (FOI)
- Filter the input signal with a Notch filter around the FOI
- Subtract the result from the initial signal
- The result is the "interesting part" of the signal, with mainly frequencies around the FOI
Why would one do that rather than applying directly a bandpass filter of similar order and bandwidth ?