This article presents a deep-learning-based approach to quantitatively extract misalignment information in the orbital angular momentum (OAM) radio communication systems. It is shown that the misaligned field exhibits unique sequence-based features for different misalignment information. Consequently, recurrent neural network, as a powerful method to deal with sequential data, is introduced to extract its intrinsic features and estimate the misalignment values. To further reduce the total number of parameters as well as improve the performance, a gated recurrent unit is employed to construct the recurrent layer. As far as we know, this is the first use of deep learning for the estimation of the OAM misalignment information in radio frequency. The setup of the training data and the performance of the proposed model under four different misalignment cases are demonstrated. To further validate the reliability of the proposed models, the trained model is evaluated on the sample carrying 20% Gaussian noise. Further, the effects of the position deviation and the number of samples are also investigated. Besides, the comparison with other methods are presented. It is illustrated that the tailored model can reach the same measurement resolution as the previous literatures while requiring only one-eighth of the sampling points.