Skip to content

Latest commit

 

History

History
66 lines (59 loc) · 2.86 KB

NOTES.org

File metadata and controls

66 lines (59 loc) · 2.86 KB

Various NOTES

filter.rs

Various normalization schemes

  • Adding a constant mean to initial stars to make always positive magnitude does not change the final result
    • B/c we do a mean subtraction from the final star in analysis thus this is moot
    • let stars = AF::add(&stars, &10.0f32, false);
      • Before stars_dc_removal and other formatting is done
  • log10 the star before analysis does not really change result
    • Probably due to how log affects fft output
    • let stars = AF::log10(&stars);
      • Before stars_dc_removal and other formatting is done
  • let stars = AF::exp(&stars); same as log10
  • more investigation would need to performed on low-pass filtering
    • some issues might be…
      • already low frequency signals
      • how does this benefit the fft performance
      • can this be done to templates beforehand
      • is this done already by the window functions
    • let b_coeffs = AF_Array::new( &[1.0], AF_Dim4::new(&[1, 1, 1, 1]) ); let a_coeffs = AF_Array::new( &[1.0 ,6.39245322,20.43172909,42.80206107,64.88239627,74.23342926, 64.88239627,42.80206107,20.43172909, 6.39245322, 1.0], AF_Dim4::new(&[11, 1, 1, 1]) ); */ //let stars = AF::iir(&b_coeffs, &a_coeffs, &stars);
    • in same place as log10

DC Removal

It appears I had some issues with DC removal logic (in the maths). So initially I removed all DC b/c it seemed to make any filter work. Now I need to investigate adding them all back. It may be that the best scheme is to only remove the DC from the templates and not the stars. As put in another note we were having an issues with weird output (esp. with abs)

\\ ------/ ------

resulting in the following image like phenomeon occuring when it should only be increasing. Focusing only on reals seemed to solve it partially but still gave strange results in lack of amplitude and for the NFD test dataset. Thus, through some testing we removed dc removal from the star windows and that seems to have fixed the issue along with (maybe boosting detection level – need to validate). Mismatched filters still seem to pretty much ignore the pressence thus with further validation this will probably be the best way going forward.

  • [ ] TODO further validate these claims…
    • mismatch effect (make it less than with or without)
      • ie with dc is the mismatch amplified or diminished and vice-versa
    • check removal of template
      • both with
      • and singularly opposed to (turn back on star dc removal)
    • detection rate of NFD and generated dataset
    • the addition of a DC component to the result and does this change based on overall DC component of each signal
      • this does appear to be the case X/
        • look into adding some kind of average DC subtraction logic for stars
  • [ ] TODO add back other code and notes from diff