In one of our designs we are doing signal conditioning of a load cell (Wheatstone bridge)
We have an issue where the main instrumentation amplifier an AD8224 has an output signal that wanders by around 20mV over the course of 10's of seconds.
We have gone through troubleshooting the signal chain and tracked the issue back to here, the above plot is taken with the inputs to the amplifier shorted out and connected to 0V.
Supply to the op-amp is +-15V, and has 0.1uF bypass directly near the IC, and 10uF about 3cm away. We are using a gain resistor of 49.9ohm which should give a gain of around 991.
We are not worried about the high frequency noise, this gets tidied up latter in the chain.
I was able to find one of our old board revisions that does not show the same issue, these boards are modular, and I was able to slot in the new and old next to each other to compare the results. The old board used a TI INA2126 instead of the AD8224.
Channel 0 red is the old revision and shows what I would expect of changes due to thermal effects.
Channel 1 yellow shows the same board as per above.
In terms of ruling out other factors, this is in a room without much airflow, breathing on the board doesn't seem to have a large effect, nor does insulating it from airflow by wrapping it in rags. My reference pins are tied directly to zero volts. I have confirmed I get a nice clean (1-2 bits of noise) signal recorded if I short the output of the amplfier to 0V (as expected).
I am all out of ideas, is this expected performance of the AD8224 (from the datasheet I think I should be getting at least an order of magnitude better).
I am certainly not an expert, and haven't had to do much careful temperature control on a project before, besides dealing with power dissipation and temperature co-efficients of shunt resistors.
I am just trying to figure out what I can do to try and replicate the plot found in the datasheet "Figure 10. 0.1 Hz to 10 Hz RTI Voltage Noise (G = 1000)" I would have expected to be able to get something in the same ballpark but since I am measuring at the output with the scale in mV because of the 1000 gain.
I have been able to replicate the results across at least 5 boards, and at least one of the boards was assembled with ICs from a different shipment.
Did the layout look reasonable?
I've asked around for some help with your issue.
Your layout looks ok to me. I'm assuming the traces that run off the picture to the right (AD8224 pins 6, and 7 for REF1 and REF2) go down to via ground close-by. With C1-C6 replaced with shorts to ground as you have it, I'd imagine it'd not be necessary to cut the traces to +IN and -IN at pins 1, 4, 9, and 12 of AD8224. However, if I were you I'd try cutting those traces just to make sure nothing is coupling in from the these connections (although it'd be hard to believe that anything would if you've already grounded these - but just in case).
I'll keep you posted if something comes up whilst discussing with colleagues.
For the level of precision you are trying to achieve, I wouldn't even consider a two layer board.
Four minimum. I would also not use a dual. Interaction, cross talk, layout problems, etc.
For figure 10, they probably had a low pass filter with a corner of 10-20 Hz.
I'm assuming your scale signal is not that fast, so I would consider slower, autozero InAmps.
LT2053, AD8421, AD8237, LT1167
or, to solve the whole problem and reduce system cost, look at the ADA4558.
Any solder joint has a tempco of 5-35uV/C, so balanced layout is extremely important.
My colleague has used an EVAL board to see if we can duplicate your issue? We configured the first (Dual) site so that one of the in-amps has a gain of 991, using +/-15V supplies, starred the inputs to the Ref gnd, and then measured at the output of the channel with the high gain. The attached scope photo shows the noise over ~400s. At this scale (10mV/div) no "slow" wander is visible to me, unlike the scope photos you'd attached.
We did notice that the measurement was very susceptible to external noise; moving a cell phone near the board on the lab bench would result in large noise spikes seen on the scope. So, perhaps some shielding would help you.
Please take a look at Harry's comments above as well.