Vortex beams carrying orbital angular momentum can be used in free space optical communication and have greater data coding freedom. However, the atmospheric turbulence channel is the key factor that limits the performance of vortex free space optical communication. It carries different random disturbances of the Laguerre Gaussian beam wavefront phase with orbital angular momentum modes, so it cannot keep its original orthogonality; it may cause serious signal crosstalk and receive light signal phases and intensities of ups and downs. When the vortex beams carrying information are transmitted in the atmospheric channel, it will be disturbed by turbulence and cause phase distortion. Vortex beam orbital angular momentum detection methods mainly include optical methods and machine learning methods. However, the optical method needs to use more communication system elements, the operation is complicated and the detection of vortex beams with phase disturbance is not involved. At the same time, the current machine learning methods have not involved the detection of high range orbital angular momentum modes so far, and need to know the prior knowledge of atmospheric turbulence. Aiming at the problem of measuring the orbital angular momentum mode information of vortex beams after phase distortion, this paper proposes a method to detect the orbital angular momentum mode information of vortex beams disturbed by atmospheric turbulence by using the designed convolutional neural network. A turbulence model of anisotropic atmospheric turbulence is established based on the multiphase screening method. Laguerre Gaussian beams with different beam parameters are transmitted in the simulated turbulence channel, and the intensity image of the Laguerre Gaussian vortex beam with phase distortion is obtained at the receiver end. In addition, the size of the receiver plate is fixed, and different sizes of the receiver plate are used to receive vortex beams with different ranges of orbital angular momentum. Due to the randomness of atmospheric turbulence, 10 vortex light intensity images under atmospheric turbulence were randomly simulated as data sets to make the model universally adaptable. To achieve better recognition performance, a convolutional neural network structure was optimized, and the orbital angular momentum recognition model based on a convolutional neural network was designed with 12 layers. In this method, the phase-distorted Laguerre Gaussian vortex beam intensity image is input as sample data. The network uses the input data set for independent learning, and after multiple iterations, it can accurately detect the high range of orbital angular momentum information of the beam. The simulation results show that when the intensity of atmospheric turbulence is uncertain, the accuracy of the model for the detection of vortex beams with orbital angular momentum modes of 1~40, 1~100, and 1~160 is as high as 94%, 90%, and 86%, respectively. For the vortex beam with orbital angular momentum mode 100, the accuracy of detection after 1 km, 2 km, 3 km and 4 km reaches 90%, 80%, 79% and 78%, respectively. When the radial index p is 0, 1, 2 and 3, the recognition accuracy is 90%, 85%, 80% and 79%, respectively. When the waist radius was 0.02, 0.03, 0.04 and 0.05, the model recognition accuracy was 90%, 87%, 83% and 82%, respectively. When the wavelength is 1550 nm, 1310 nm and 850 nm, the recognition accuracy is 90%, 88% and 86%, respectively. The proposed method has high accuracy and learning efficiency, and there is no need to convert the scroll beam into a beacon beam before detection, which reduces the number of components in the communication system, greatly reduces the complexity of the system and improves the detection rate of the scroll beam. The research results are of great significance for the application of scroll optical communication systems.