It seems that only the orientation is not the desired one. Hence you get correct values but they are not the one you expected.
To make it systematic I would propose you to build the transformation matrix between your desired sensors location and the one what was used in the survey. This way you get a framework also working if the position of the sensor is off.
see http://www.diag.uniroma1.it/~deluca/rob1_en/08_EulerRPYHomogeneous.pdf for some theory
Then you can build the transformation matrix of your sensor reading, aka a rotation matrix construct from the angles, then by multiplying the two transformation matrix you wold get the transformation matrix of your measurements at the give point. It is up to you to go back to roll, pitch, yaw angles. Note that RPY are ambiguous and subject to gimbals lock phenomena, hence rotation matrices are preferred even if they are less compact.
A comprehensive guide to go from one orientation representation to the next, https://www.swarthmore.edu/NatSci/mzucker1/papers/diebel2006attitude.pdf