The utilization of beam-carrying orbital angular momentum (OAM) for free-space optical (FSO) communication can increase channel capacity. However, the misalignment of the beam is an effect that must be mitigated in FSO communication systems. Due to the robustness of deep learning technology in pattern recognition, a neural network structure is proposed and improved to mitigate the effect of misalignment error. First, compared with the simple convolutional neural network proposed, data augmentation is adopted in the training. Then, a view-pooling layer is added after the convolutional layer. This layer can longitudinally compress feature maps from multiple receiving angles. In order to verify the performance of the proposed method, related experiments are reported in this paper. It can be seen from the results that when the tilt angle is less than 35°, the accuracy of OAM mode detection is above 99%, 93%, and 88%, respectively, corresponding to the condition of weak (Cn2=1×10-15 m-2/3), medium (Cn2=1×10-14 m-2/3) and strong (Cn2=1×10-13 m-2/3) turbulence.