In my opinion, we can try to improve performance by not calculate the acceleration norm for each of the four sensors but keep the accelerations for each axis.

Because with the norm, we lose the direction that could be useful to improve the performance.

Another thing you could try is to use a neural network approach.