On the process of alignment for pupil off-axis telescope system, the determination of decentering and tilt misalignments is a key step. In this paper, we proposed a new method which adopts the fully connected neural network (FCNN) as a fitting tool to establish the nonlinear mapping relation between misalignments of different fields of view (FOVs) and Zernike coefficients. Firstly, we establish a pupil off-axis reflection telescope model, then decentering and tilt misalignments are introduced to acquire corresponding aberrations that represented by Zernike coefficients. We use aberrations as the inputs of FCNN; misalignments and FOV as outputs. FCNN is trained by the combination of inputs and outputs as a dataset, and we use a new dataset to test the accuracy and effectiveness of the trained FCNN. The results show that the mean absolute error (MAE) of the X-axis decentering error, V-axis tilt error and angle of view are 0.0506 mm, 0.0204 degrees and 0.0124 degrees, respectively. These results demonstrate that the proposed method is effective and feasible to calculate the misalignments.