Accelerometer calibration: why it does not work?

There is a perfectly simple method for calibrating accelerometer offset and sensitivity using 4-points rotation. The advantage is this method does not depends on the angle between sensor axis and g-axis. The only problem with it is that it does not work. Equations look correct and quite obvious. Does anybody have an idea what is wrong in this method? Details: I simplified calibrating device. Instead of calibrated disks etc. I took a wooden bar with 2 inches sides, 20 inches long, attached metal angle to one side, bent it so that angle's surface is at approx. 30-40degrees related to bar surface and mounted sensor on metal angle surface. Then I took sensor readings (averaged over 1000 values) when turning the bar so that it stands on each of four sides, so this guarantees that measurements were taken on 4 'equally spaced angular positions' with 90 degrees difference. Readings were stable and permanent, and reproducible for each of four sides. I calculated offset values using formula from the above source. The problem is when I re-mounted sensor under a different angle and repeated measurements, I get completely different values for offset and sensitivity.

So there is a flaw somewhere in this approach; besides it looks very obvious and simple, but I never met this method described anywhere else, which makes me to suspect there are some reasons why it is not used. Any ideas?

"The axis of sensitivity is in a vertical plane (within a few degrees)."

Was you accelerometer axis of sensitivity perpendicular to the wooden bar you mounted it on?

Obviously not, as I understand it, those formula should work whith any angle? It is saying there 'The method presented here does not require any alignment with gravity' That moment is somewhat unclear to me. What actually means 'axis of sensitivity'? Sensor has 3 axis, and it seems that none of them is required to be aligned with gravity, . Can you please explain it in more details? Oh, I am starting to understand. Rotation effectively eliminates misalignment in vertical plane only, or, rather, in a plane orthogonal to rotation axis :( So to make it work for any angle, I must mount it on a cube and take 12 readings. Thanks for a hint, johnwasser!

The method obviously cannot work if the horizontal axis of rotation and the sensitivity axis of the accelerometer being measured are parallel. The accelerometer would always indicate 0g so the sensitivity cannot be measured.

You can do a 2-axis accelerometer with four positions. If the plane of the accelerometer is perpendicular to the wooden bar you only need to take readings with the bar on each of the four long sides (readings from each axis). If you have a 3-axis accelerometer you need two more positions.

I am having the same problem with calibrating my sensors. Thus far, I have attributed it to variances in electrical operation at device start-up. But recently I have been thinking of using some trigonometry and real physics for calibration. I will try this experiment out soon!

Thank you for the informative post. :) I'll post again once I have results.

Btw, If ever anyone views this (as I see its an old thread) might I know about your results and your own methods for calibration? Any and all feedback is appreciated.