Orbital Angular Momentum (OAM) beams, with their theoretically infinite topological charge and orthogonality, demonstrate significant potential for enhancing the capacity of communication systems. In OAM Free-Space Optical (FSO) communication systems, atmospheric turbulence acts as a significant technical bottleneck, causing optical wavefront distortion and leading to effects such as scintillation, beam expansion, beam wander, and arrival angle fluctuation. These effects reduce the recognition accuracy of OAM beam topological charges, thereby increasing the bit error rate in OAM-FSO systems. To address this, a deep learning-based approach utilizing Convolutional Neural Networks (CNNs) has emerged, capturing OAM beam intensity images after turbulence transmission via Charge-Coupled Devices (CCDs). However, the recognition speed is constrained by the low frame rate of area array CCDs and the relatively slow processing speed of deep learning on PC. CCD cameras typically operate at tens of Hz, while PC-side data processing rates are generally in the kHz range, which is insufficient for high-speed OAM-FSO communication applications.This paper proposes a novel recognition scheme to improve CNN recognition speed. The scheme employs a cylindrical lens and Photodiode Array (PDA) to convert the two-dimensional intensity signal of superimposed OAM light fields into a one-dimensional signal, thereby reducing the computational complexity. A one-dimensional CNN is implemented on a Field-Programmable Gate Array (FPGA), referred to as CNN-FPGA, enabling high-speed signal processing and recognition. The scheme is theoretically analyzed using the Hill-Andrews model and random phase screen simulations. The same dataset is used to train and test both the refined CNN, referred to as model2, and a conventional CNN. The model2 achieves 99.0% recognition accuracy in a 200-m channel with a turbulence intensity of Cn2=5×10-15m-2/3, requiring only 576 floating-point operations (Flops), while the conventional CNN achieves 99.9% accuracy under the same conditions but with 67 520 Flops. Although the refined CNN's accuracy is slightly lower, it remains within the effective recognition range, and lower bit error rates can be achieved using coding techniques. To evaluate the adaptability of the trained model2 to atmospheric turbulence disturbances of varying intensities, the recognition accuracy of model2 for OAM optical signals under different turbulence conditions was simulated using six training parameters. The simulation results demonstrate that the trained model2 exhibits strong adaptability to atmospheric turbulence disturbances across a range of intensities. In the experimental process, the superimposed OAM beams with topological charges l=±1,±2,±3,±4 are generated using digital holography. To determine the turbulence intensity during the CNN-based recognition of superimposed OAM optical signals in the experiment and compare it with simulation results, the atmospheric turbulence intensity at the experimental site was measured using the angle-of-arrival fluctuation method. The OAM transmission recognition experiments conducted in a 200-m outdoor environment, approximately 3 h after sunset with an atmospheric refractive index structure constant Cn2 on the order of 10-17m-2/3, demonstrated a recognition accuracy exceeding 93.5%. Compared with simulation results, both the experiment and simulation exhibit a consistent trend of decreasing recognition accuracy with increasing turbulence intensity. The CNN-FPGA exhibits a maximum logic delay of 5.4 ns and a net delay of 12.6 ns within the FPGA chip, resulting in a shortest clock period of 0.018 μs. Compared to CPU and GPU platforms, the FPGA achieves comparable recognition accuracy while offering the fastest inference speed, the lowest cost and power consumption, and a more compact design with enhanced portability. The proposed scheme integrates a high-speed PDA module, optoelectronic signal processing circuit, analog-to-digital converter, and FPGA-based CNN module, achieving a recognition time of 0.072 μs per OAM signal, which greatly improve the recognition rate of optical communication systems. This scheme provides an efficient and practical technical solution for OAM optical signal recognition in FSO communication systems, which holds reference significance for the development and application of OAM-FSO communication technology.