Re: High Resolution AD Converter Noise

In article <42add359$0$27794$9b4e6d93@xxxxxxxxxxxxxxxxxxxxxxxxxx>,
janvi@xxxxxxxxxxxxx says...
> A customer is using a ready off the shelf AD Converter board using 8
> differential channels with a resolution of 16 bit @ 100 khz. The circuit
> contains 8 parallel switched capacitor SAR (no analog mux) from Analog
> Devices AD976. It is using the internal reference and the input circuit is
> a INA103 from Burr Brown. With gain=1 the input range is +-10Volt.

If the board is off-the-shelf, please specify which board it is and
whether the schematic diagram is available on line. In the best case,
there may be someone here who has experience with that particular board.
> Now customer complains about unknown noisy results (of about 20 LSB) using
> that board. To examine this, I use the board at my table in clean
> environment. After connecting all differential ends to ground, I can easily
> see 10 diffrent numbers at a simple 100mS refreshed display what is equal
> to an error of 10 LSB. With a gain of 1, the INA103 can hardly cause this
> noise.
> My last AD design was long ago with only 12 bit resolution at lower speed
> but I was able to see with a 10gear poti any number at the screen without
> its neighbour and it was even possible to adjust the poti to the edges
> where the result toggled to the upper or to the lower neighbour.
> I can hardly believe the AD976 is poor becouse the noise is much above the
> INL what is specified with +-2LSB max. I have no idea how the given
> "transition noise" of 1 LSB is defined in the data sheet. The gausian
> histogramm shows only 7 codes for DC code uncertainty. 4 of them are
> extremly rare and only 2 of them make about 2/3 of the results.
> Assume all, I think the board layout or grounds are poor and we meassure
> some crosstalk of the bus or the close DC/DC converters. Therefore it would
> be nice to hear your experience with AD convertes at similar resolution and
> maybe someone already examined a 16 bit or above reference board with its
> own design?

The DC-DC converters can be a problem---especially if they are running
their oscillators at a few hundred KHz and you are collecting data
at the same rate.

If you don't need 100KHz sampling rates, you could oversample and
use the average to reduce noise. This will work if you are seeing
noise with a Gaussian distribution. If the noise is nearly synchronous
due to the DC-DC converters, oversampling may not work as well.

I've used the AD974 (AD976 with 4x front end mux) and gotten about
2 LSBs noise (standard deviation of 1000 samples) with a nice
Gaussian distribution of the samples.

Mark Borgerson