I've collected accelerometer+gyro data on people executing some actions using a sensor that is attached to wrists/ankles. I'm planning to train a neural network with the data.
There wasn't any supervision in how the sensor was mounted so there is a reasonably large variance in the orientation of the sensor. I am considering using data augmentation to account for this (i.e. train the network to deal with any orientation). The different orientations are just a rotation around the bone so it is physically valid.
For the accelerometer, this is quite straightforward, I just need to compute a quaternion that transforms the accelerometer data according to 3 augmentation rotation angles.
I'm not so sure about the gyro data. The sensor outputs angular velocity in 3 axes which can be put into a quaternion, $[1,dt*gx,dt*gy,dt*gz]$ where dt is the sampling timestep and gx/y/z are the gyro axes values.
Can I just rotate the gyro quaternion using the same quaternion I used to augment the accelerometer data?