I need to track an object in 3D-space and I can use only accelerometer and gyroscope. I can't use a magnetometer because the object itself has a strong magnet.
I need to know the orientation of the object (how much gravity there is in $x, y,$ and $z$ components), and how much acceleration there is in $x, y,$ and $z$. I'm not trying to calculate the object's absolute position.
Is it possible to create sensor fusion that has less than $10\%$ error in gravity and acceleration? At this point I'm just trying to understand if this goal is realistic.
I did a quick experiment with my smartphone and found out that the way it calculates gravity (LPF?) results $\approx 40\%$ error when phone is moved.
This was just a test: I'm not making phone app. I'm designing dedicated HW.
I'm just posting here a pic to show how the "gravity sum" looks like after fixing the math as suggested by Maximilian.