Here is how I did in my system (AutoBuddy)
First I “calibrate” the tag. I use samples (200 in my case) and I average the value on each axis. I then use that as the “zero” value and subtract it from subsequent measurements to normalize them.
After that I detect movement using the “vector” (i.e. sqrt(x^2 + y^2 + z^2) with x,y,z normalized) If I get a vector > 200 I detect movement.
You could bypass the normalization step and compare the vector value to 1000 (milli g) which should be the vector value at rest.
Hope this help.