Mechanically damping your IMU may be the answer, but it can be difficult. If you use foam it'll be hard to maintain calibration. If you use a servo-mount style with grommets and spacers you may not damp out the blade-pass frequency. Basically you have to make sure you have enough mechanical travel in your soft mount to absorb the accelerations of your helicopter fusalage.
You may also be able to get better performance by going to the ADXL210 (see my comments on saturation below), or a combination of the ADXL210 and some mechanical damping.
Using a speed-calibrated anti-bias is an idea, but it'll be very sensitive to the balance of the blades, and it'll probably not hold over temperature variations or changes in the individual part. If you're desperate for a prototype, if your situation is such that the ambient temperature won't vary much and if you don't mind recalibrating any time you change blades go ahead.
The loudspeaker test is a very good idea, by the way.
------------
I suspect that you are seeing one of three effects, none of which are entirely consistent with what you mentioned. Here's what I can think of:
- The part works by vibrating a little micromachined member, and demodulating what it sees. If you vibrate it at a frequency that's harmonically related to it's oscillator then it'll show bias.
- The part has a +/- 2g range, and it works by measuring acceleration, _then_ filtering the signal. Your helicoptor vibration will be much greater than 2g's. If you're saturating the internal amplifiers or the mechanical travel of the MEMS device it may hit one rail before the other -- this will cause an asymmetry in the internal signals, which will filter out to a bias.
- The part may be sensitive to cross-axis acceleration, which may resolve to a bias. If this is along the g vector it'll seriously ease the problem of building a good soft mount.
And here are some questions to ask of your system, to resolve if the trouble is one of the possibilities that I've mentioned or something else.
In your loudspeaker tests, which axis was the part most sensitive to? You should do the test in each of three directions. On-axis would indicate simple saturation, cross-axis would indicate that something more complicated is going on.
Was there any sign of thresholding in your loudspeaker tests? If you were to plot the bias vs. signal level is there a portion of the plot where the bias is zero for small signals with a reverse knee where the bias suddenly becomes proportional to level? If so, you're saturating, and the ADXL210 is probably the way to go.