I'm not sure I'd go w/ the temp gradient as a first issue. First issue is 2 wire gives much poorer accuracy and isn't officially supported. We have 2 wire RTDs in aerospace application simply because designers didn't want the extra weight of more conductors. For test sets I'll always go w/ 3 or 4 wire.
If the RTD is in contact w/ chassis (film RTD on surface or a probe that is in contact w/ the probe sheath, in contact w/ chassis or plumbing) then ground noise voltage will create common mode (noise on both wires). I've seen inverter drives often have capacitive coupling into chassis that can mess up all kinds of readings as soon as you energize the inverter. Likewise they can radiate and if you don't have a shield properly drained on one end only, a 1000 ohm RTD is more vulnerable than 100 ohm since it's higher impedance. The filters on the 9226 are really for 60 Hz noise and are not as good in high speed mode if you are using it.
If it is common mode, you can look at each sense line with a scope and see the same noise signal appear when the noise source turns on. If you have multiple inputs, I'd disconnect one at a time and see if only one is the issue. It's a multiplexed A/D so noise on one can bleed onto the others, in my experience w/ similar modules (haven't used the 9226 itself). And last, yeah, capacitors across the sense lines will help normal mode noise. If you know the noise source frequency, you can tailor the cap value to make a low pass filter that addresses the frequency but minimizes impact on the real signal. Higher frequency common mode noise can be reduced with ferrite clamps, too.
