The x,y,z acceleration values in your sample file are highly correlated on time, so using delta values for them should also improve their compressibility.
Then, you can divide the x,y,z deltas by the time deltas in order to get the derivate of the acceleration vector in order to get even more normalized values. The problem is that at this point you are using floats instead of integers, but unless you want absolute precision this should be a minor issue. Just select a resolution that fits your problem and discretize them accordingly.
Another possibility is to express the x,y,z deltas (or derivates) in the coordinate system defined by the Frenet-Serret frame. That would probably produce numbers that are nearer to zero and so, more compressible.