Hi, I'm not familiar with k-Means but I have some tips that you could use: The Sound and vibration toolset and the Signal processing toolset offer a Zoom FFT tool that you can use to extract more detailed information on frequency range with the same number of points. It uses a donw-conversion algorithm to step down the data and the perform an FFT. a better prescition on the peaks can help you classif better into families.
I'm may be completely showing my ignorance of you application here, but you could create an array of the peaks and perform a histogram where your main peaks could determine the center of each bin in the histogram. Another idea would be to do some averaging over your FFT data to create a smoother curve with peaks where the families would be. Just my 2 cents! Juan Carlos N.I. FYI: LabVIEW's Order Analysis Toolset has tools for order tracking and machine monitoring.