The structure and measurement principle of common magnetic dip meter are introduced in this paper, whose possible error and its source are analyzed. For the azimuth, one of the main parameters of borehole attitude measurement, a three layers RBF network is established based on RBF neural network compensation algorithm, which input is a two dimensional vector which is made of measured deviation angle and azimuth, and output is the expected azimuth, and the sampling data of the magnetic dip meter are used to test. The experiment results show that, the modeling time of the RBF neural network compensate algorithm is short, and the azimuth precision can be improved from±2.1°to ±1.9°or better.