You need to describe your system as a linear system, preferably with a matrix notation.
then change the base or make the state variables exactly what you want to measure, finally with an ¿observer' you would estimate any state if it is observable even if you are not measuring it directly, if it could be represented as a linear combination of its states then it will be observable.
Using two giroscopies in a rigid body actually does'nt help too much, just reduce a little the radius of the gaussian noise (the variance of normal distribution is the mean of the variances, probably reduce noise in half). It could be different if they are both rotating since the correlation will be broken and the information provided by them won't be the same/similar.
But you probably don't need two giroscopies if they are attached to the rigid body...
To achieve what you want neither need sensor fusion, you only need to describe the system with inputs motor velocities and output as the angle of the rigid body, then close the loop using one gyroscope in order to feedback the error.