Needing more than a spark test?

[graham-xrf]

Thing is, even if it looks like it wouldn't take much of a gradient to get a 500uv difference between 6.5mv/K and 4mv/K, the interfaces are VERY close together -- think IC pin to solder, then the same solder pool to the copper PCB trace right underneath it. Not much separation, physically and thermally speaking. So while some offset is theoretically possible, I don't think that is what is going on here.

I agree. The dissimilar metals voltage, even if it were a very significant big voltage, connects both metals to a journey through copper, and then goes through another soldering of the same metals, completely cancelling itself if the temperature at both ends is the same. They are all on the same board, and there is no raging dissipation to make one end way hotter than the other.

The voltage is tiny to begin with, even before it cancels. If the temperatures both ends are equal, then I think offset is not even theoretically possible!
This time, I would say thermocouple voltages are unlikely to be the cause.

 

[WobblyHand]

Well, since it's not that hard to do, I will remove the part and clean off all the solder and possible flux and try again. Sometimes it is the nut behind the wheel. And sometimes it is the board. Can't do much about the board, but I can try redoing the attachment. If that doesn't work, I will replace the part. If it fixes it, great. If not, back to the drawing board.

Edit: removed part, found a little flux underneath. Cleaned and put back the original part. Set up routine to continuously output mean count and RMS variation about the mean. Basically no difference, but it is easier to see cause and effect now. Collect 16K samples, compute the above and print out, wait 500ms and continue. The ground node of C11 is sensitive to touching it with a metal tweezers. It drops a count from 14 to 13. Touching the filtered side drops the count average to 6, but the RMS noise jumps up, becoming unusable. That's because I'm injecting stuff into the circuit. Touching the filtered trace from C11 to the ADC input also changes the average count, even though the trace is covered. Seems pretty sensitive to the capacitance of the trace. Also that trace is quite long. Touching either side of R2 (the 15 ohm) resistor makes the average count drop but dramatically increases the RMS variation. The filter pole is about 4 MHz. I can try replacing the part, but touching the ground of C11 and seeing stuff change is telling me something.

 

[homebrewed]

From Analog.com 404 "page not found" crash, there is a temperature sensor (Heath-Robinson type) feeding back into a Analog "404" quad flatpack, looking like it is doing something a little similar to Figure 30 on Page 20 of the AD7667 datasheet.

View attachment 437645

A shift in Vref due to temperature would produce a shift in the ADC's scale factor, which will only indirectly affect an input offset voltage. It won't cause offset. Looking at Figure 21 in the DS, the worst-case unit's Vref change from 20C to 40C is about 100uV out of 2.4984V. That's less than the specified 3PPM/C, which would produce a shift of about 150uV.

Looking at it from the perspective of a ~13-count offset, a 100uV shift in Vref would result in an apparent offset shift of around 20nV for a 20C temperature change. I calculated this using: 13*(Vref1/65535 - Vref2/65535)

It seems to me that ADI's practice of equating a shift in Vref to being being equivalent to "X" number of LSB's is misleading. What it WOULD do is produce an error of "X" counts for a full-scale input. In practice that error would be spread out over all 16 bits.

 

[graham-xrf]

It seems to me that ADI's practice of equating a shift in Vref to being being equivalent to "X" number of LSB's is misleading. What it WOULD do is produce an error of "X" counts for a full-scale input. In practice that error would be spread out over all 16 bits.

Yes - I don't like that either. I think it is called "specsmanship", and I suppose that is a fairly extreme example of choosing the baseline to make the numbers "look better".

I suppose one could use pin 45 (TEMP) to contrive a little linear analogue to feed into the scheme to compensate the effects of temperature. Given the results of your calculation, that is, I think, ultra-nerd OTT unnecessary, and that is coming from one like me!

If, later, in the unlikely event we discover what we have is too touchy over temperature gradients, we can figure out something. That said, I sure would like to nail down what the large(ish) offset Bruce is discovering is really all about.

 

[graham-xrf]

Getting back to pulse magnetics
Not really a déjà vu from better times. There was a time when I would happily delve the black arts of switcher PSU inductor and transformer design. I even have (somewhere) my kit of non-saturating 2-section pot cores and bobbins, in the various sizes. I don't really want to go back there, tangling with AL parameter, and leakage inductance, and suchlike.

The little 1:1.3 transformer for my ultra low noise PSU can be a HALO brand thing, available from Mouser at £10 for one! That may even be fair, but I was hoping for something nearer £1.50, or £2.10. Hmm.. maybe Murata..

 

[homebrewed]

Well, since it's not that hard to do, I will remove the part and clean off all the solder and possible flux and try again. Sometimes it is the nut behind the wheel. And sometimes it is the board. Can't do much about the board, but I can try redoing the attachment. If that doesn't work, I will replace the part. If it fixes it, great. If not, back to the drawing board.

Edit: removed part, found a little flux underneath. Cleaned and put back the original part. Set up routine to continuously output mean count and RMS variation about the mean. Basically no difference, but it is easier to see cause and effect now. Collect 16K samples, compute the above and print out, wait 500ms and continue. The ground node of C11 is sensitive to touching it with a metal tweezers. It drops a count from 14 to 13. Touching the filtered side drops the count average to 6, but the RMS noise jumps up, becoming unusable. That's because I'm injecting stuff into the circuit. Touching the filtered trace from C11 to the ADC input also changes the average count, even though the trace is covered. Seems pretty sensitive to the capacitance of the trace. Also that trace is quite long. Touching either side of R2 (the 15 ohm) resistor makes the average count drop but dramatically increases the RMS variation. The filter pole is about 4 MHz. I can try replacing the part, but touching the ground of C11 and seeing stuff change is telling me something.

I think I know what may be going on. More-careful examination of the AD8655's DS shows that its Vol does NOT NECESSARILY include ground. It's specified to get CLOSE to ground (apparently close enough to be called "rail to rail"), but in fact its Vol is allowed to range from the specified typical 8mV up to 30mV max. If the input stage _is_ capable of operating down to ground, the amplifier will be open-loop at that point so all bets are off when it comes to gain accuracy, offset etc.

This may well be a common problem for so-called RR amplifiers. I tried using a different one in an application and it never worked right, turns out that it didn't get close enough to the rails for my requirement even though the mfr called it a "rail to rail" amplifier.

The other tricky thing about rail to rail amplifiers is that they can come in different flavors. Some have rail-rail inputs but their outputs aren't. Or vice-versa. Since we need an amplifier whose input and output include ground, we need a true RRIO device.

The interesting thing is that the amplifier ADI recommends for feeding the AD7667 is NOT a rail-to-rail amplifier. Also its minimum supply voltage is +/-5V.

Some surgery and an external power supply will tell the tale. But, since the AD8655's maximum supply voltage is only 5.5V, the negative voltage can't be TOO negative. To ease things up, the next version of ADC board may run the AD8655 at +3.3 and -2 (or thereabouts).

Are you or Graham aware of any amplifiers that might be suitable drop-in substitutes? It may be difficult to find a drop-in amplifier that meets all the requirements AND happens to be in stock..

 

[WobblyHand]

I think I will remove R2, isolating the AD8655, and grounding C11, the input to the ADC. Unfortunately, at least from the schematic I have, it is difficult to discern which ground we are talking about. In my mind, the buffer amp ground should be connected to AGND of the ADC. I might try just jumpering pin 42 to pin 43, and removing R2 and C11, at least to see the offset go away.

@homebrewed good catch on the AD8655. Rail to rail, indeed...

 

[homebrewed]

Getting back to pulse magnetics
Not really a déjà vu from better times. There was a time when I would happily delve the black arts of switcher PSU inductor and transformer design. I even have (somewhere) my kit of non-saturating 2-section pot cores and bobbins, in the various sizes. I don't really want to go back there, tangling with AL parameter, and leakage inductance, and suchlike.

The little 1:1.3 transformer for my ultra low noise PSU can be a HALO brand thing, available from Mouser at £10 for one! That may even be fair, but I was hoping for something nearer £1.50, or £2.10. Hmm.. maybe Murata..

Your low-noise, low-cost PSU may come in handy, if my theory turns out to be correct! Do you have a P/N for the particular transformer you're talking about?

 

[WobblyHand]

Your low-noise, low-cost PSU may come in handy, if my theory turns out to be correct! Do you have a P/N for the particular transformer you're talking about?

Look for something center tapped to make + and - rails. Single ended stuff doesn't seem to work so well down near ground, as we are finding...

Turns out that the AD8655 has a typical 8mV V_OL at 1mA with a 5V rail. The fact I got 0.58 mV and it wanders is telling me this isn't a good long term set up. I'm probably lucky to get this low of an offset, and you are even luckier!

 

[homebrewed]

I think I will remove R2, isolating the AD8655, and grounding C11, the input to the ADC. Unfortunately, at least from the schematic I have, it is difficult to discern which ground we are talking about. In my mind, the buffer amp ground should be connected to AGND of the ADC. I might try just jumpering pin 42 to pin 43, and removing R2 and C11, at least to see the offset go away.

@homebrewed good catch on the AD8655. Rail to rail, indeed...

The buffer amp's ground is AGND.