1
$\begingroup$

I am trying to understand the choices going into the digital filter design in the TI ADS1119: They give the filter response on page 18 of the datasheet, here is e.g. the curve for the 20 sample/s case:

enter image description here

The parts I think I (mostly) understand:

  • There are two "notch filters" at 50Hz & 60Hz to filter out line noise
  • The filter appears to have 22 coefficients, which together with the 400Hz sampling rate (judging by the frequency window they plot) almost results in the 20samples/s effective sample rate at the end.

The parts I don't understand:

  • Why 22 coefficients and not 20, if the goal is to neatly get to 20 samples/s at the end? (assuming you want every sample to be used once)
  • Why is the response at >60Hz not more uniform/flat? If I use e.g. the tool from here to design a filter with two "strong" (target -90dB) stop bands at 48.8-50.8Hz and 58.4-60.4Hz (tweaking the numbers until the generated filter had the bands properly centered), and a "weaker" (target -30dB) one for 80-200Hz, I get a response like this: enter image description here I can see that their response overall seems to be slightly lower, but I don't quite get why - wouldn't moving around the zeroes necessarily increase the response in some areas?

Obviously, nobody here knows what the designers were actually thinking, but I am wondering whether there are considerations that I am missing that would explain the final form of their filter.

$\endgroup$
2
  • $\begingroup$ Perhaps they are also taking a swing at the line harmonics? Line noise is rarely just the fundamental, you also want to kill 100/120 and 150/180. $\endgroup$
    – Hilmar
    Commented Jun 22 at 11:28
  • $\begingroup$ @Hilmar good point, thanks! At one point I considered something like that, but I saw that for some reason they don't hit the 180Hz peak (so I naively dismissed the idea as wrong...). But I now realize that at least the 100/120Hz peaks are properly supressed. Given that odd harmonics are usually stronger, I would have expected the 180Hz peak to be more prioritized, but I guess they had other things to consider as well $\endgroup$
    – Lukas Lang
    Commented Jun 22 at 11:52

1 Answer 1

1
$\begingroup$

The modulator sample rate is 256kHz (not 400 hz), which the digital filter acts on, so they aren't showing most of the filter frequency response. They say it settles within a single cycle, and it wouldn't make sense to discard samples, so my guess is the FIR filter has a width of 12800 samples for the 20 hz output rate case (that is, it will have more than 22 coefficients).

Regarding the stop band frequency response trend decreasing as frequency increases, I would think that is typical of most low pass filters, so I don't find such a trend surprising, especially for delta-sigma ADCs that push noise to higher frequencies. If you are referring to the unevenness in how much the gain recovers between notches, it seems mostly dependent on notch spacing, which is a consistent pattern in all the frequency responses they show, with the gain recovering less the closer they are spaced.

Finally the notches: As you say, the 50 and 60 hz notches are for filtering out mains power supply noise, with each consisting of two notches at +/-1hz of those centre frequencies to allow some uncertainty.

Now the speculative part: The 100 and 120 hz notch filters are for the second harmonics of the mains power. Beyond that I cannot see consistent patterns. For example, the 150hz may be the third 50hz harmonic, but then why isn't there a notch for the third 60 hz harmonic at 180 hz? When downsampling, it is not uncommon to have notches at frequencies that alias to 0 hz (which is a good explanation for the overall 90 hz output rate frequency response), which would explain 160hz, but then why aren't notches present at all prior multiples of 20 hz? To have the closely spaced 150, 160, and 170 hz notches makes me think something significant is happening there that could be hardware related, but then why don't the other data output rate filters have these notches? I'll end the speculation there. I cannot see a general pattern for the placement of notches that is consistent across all output data rate frequency responses.

Probably it's best to contact the manufacturer or similar if you want to know with certainty.

$\endgroup$
2
  • $\begingroup$ Thanks for your insights! Regarding the sampling rate: As I interpreted the datasheet, they are downsampling the data stream from the ADC before applying the FIR filter. Initially, I came to the conclusion that the numbers make perfect sense: "2x max frequency in the plot / number of zeroes = sample rate" (400Hz / 20 = 20Hz, 2kHz / 22 = 90Hz, ...), but the number of zeroes never quite fit... As for the height of the gain between the notches: The thigh that confuses me is that e.g. for the peak at ~85Hz, I get ~-28dB, and they get ~-33dB, even though the notch spacing is smaller in my case $\endgroup$
    – Lukas Lang
    Commented Jun 22 at 12:54
  • $\begingroup$ There's no mention that I can see of it decimating twice. The following conversation with the supplier helped give me more confidence: e2e.ti.com/support/data-converters-group/data-converters/f/… $\endgroup$
    – Stephen
    Commented Jun 22 at 13:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.