As an emerging wireless communication technology, optical communication, which utilizes lasers as a means of information transmission, combining the advantages of high communication capacity and high-speed transmission, is gradually gaining widespread attention. Since Allen and others first demonstrated that optical vortices carry Orbital Angular Momentum (OAM) under near-axial conditions, vortex beams have attracted significant attention. OAM has infinitely many orthogonal eigenstates, forming an infinite-dimensional Hilbert space. Theoretically, this property can infinitely increase communication transmission capacity. Due to these unique characteristics, vortex beams find widespread applications in optical imaging, micro-operations, and free-space optical communication.However, when vortex beams propagate in the ocean, the presence of turbulence significantly affects the quality of the beam, leading to distortion and degradation of the light field at the receiving end. In recent years, scholars worldwide have proposed OAM recognition schemes under different conditions. Nevertheless, when vortex beams propagate in complex media such as oceanic turbulence, their OAM spectrum becomes broader, posing a challenge to OAM recognition. The lack of effective OAM detection methods hampers the development of oceanic optical communication. With the rise of artificial intelligence, deep learning technologies have rapidly developed in various fields. This paper proposes a method based on improved deep separable networks (IXception) that combines deep learning with traditional oceanic optical communication technology. This method aims to achieve OAM mode recognition of vortex beams transmitted through oceanic turbulence. Initially, the paper adopts the idea of stepwise phase screens based on the power spectrum inversion method to simulate the transmission process of vortex beams with different OAM values in the ocean. The corresponding dataset is established by collecting, degrading, and distorting the speckle field image at the receiving end. Subsequently, the dataset is randomly divided into training, validation, and test sets in an 8∶1∶1 ratio, and the IXception model is trained using the training set. IXception adopts the architectural concept of Xception, combining the residual structure of ResNet and the inverted residual structure of MobileNetV2. IXception reduces the number of network parameters, network weights, and the complexity of the network model structure while improving accuracy through partial connections and weight sharing. Additionally, the network can extract highly spatially deep features, reduce redundancy in network structural parameters, and enhance generalization ability. For transmission distances of 20 m and 80 m, IXception is trained for 40 cycles using the corresponding training set, and the training and validation accuracy curves fit well, with validation accuracy reaching 99.20% and 97.9%. The results indicate that the accuracy and loss curves fit well during the training process of the training and validation sets, with no signs of overfitting, underfitting, or gradient explosion. IXception can effectively extract OAM modes from degraded light field images. In practical applications, turbulence-induced disturbances to the beam pose a significant challenge to oceanic optical communication. To investigate the generalization and robustness of the IXception model, vortex beams are transmitted in seawater at distances of 20 m, 40 m, 60 m, 80 m, and 100 m, and the corresponding degraded light field datas are collected to create a dataset. The training set from different transmission distances serves as the input for the IXception model. Four statistical measures, namely, mean absolute error (E_MAE), mean relative error (E_MRE), root mean square error (E_RMSE), and correlation coefficient (R_xy), are selected to evaluate the performance of the IXception model. The evaluation results show that as the transmission distance increases, the OAM spectrum through oceanic turbulence's phase screen also broadens, resulting in stronger distortion of the light field at the receiving end and making it more challenging to extract OAM modal values from the distorted light field. The research indicates that the IXception network architecture has strong generalization ability, even achieving an R_xy of 91.21% for OAM modes at a transmission distance of 100 m. To compare the recognition performance of the Xception and IXception models, evaluations are conducted at transmission distances of 40 m and 100 m. IXception outperforms Xception in terms of E_MRE, E_MAE, and E_RMSE evaluation results. Additionally, regarding R_xy evaluation results, Xception scores lower than IXception at both transmission distances, especially 4.66% lower at 40 m. IXception reduces the number of network weights through partial connections and weight sharing, achieving a training time 39 ms lower than Xception for the same batch (step). In conclusion, the overall performance of the proposed IXception model in recognizing OAM modes in distorted light fields caused by oceanic turbulence is superior to the Xception model. This research provides theoretical support for the practical engineering application of oceanic optical communication using vortex beams.