【摘要】 提出一种基于动态权重多任务学习的神经网络方法，用于同时对拓扑荷（l）与彗差系数（kC₃）进行识别。采用功率谱反演法构建了包含大气湍流与彗差效应的斜程传输模型，生成多种参数组合的数据集，利用ResNet18进行双任务识别。引入一种基于动态权重的训练方式，有效解决了双任务识别中的不平衡问题。实验结果表明，在近地面湍流值C₀=1.7×10^-14 m^-2/3的情况下，改变天顶角（30°、45°、60°）以及传输距离（2000 m、3000 m、4000 m），模型对拓扑荷数以及彗差系数的识别准确率均达到1。相较于平均权重，动态权重方法具有更快的收敛速度以及更短的训练时间，且性能显著优于ResNet34和VGG16网络，为解决不平衡多任务学习问题提供了新思路。更多还原

【Abstract】 Objective Vortex beams, characterized by the coupling of spin and orbital angular momentum（OAM）, exhibit a helical wavefront and a doughnut-shaped intensity profile with a dark core at the center. When propagating through atmospheric turbulence, these beams suffer from a series of detrimental effects, including intensity scintillation, phase perturbation, beam spreading, and beam wander, all of which deteriorate transmission quality and limit their performance in optical imaging and communication systems. In practical scenarios, slant-path transmission models better represent long-distance atmospheric links such as ground-to-near-space or satellite-toground communications. However, most existing deep learning-based recognition methods are primarily designed for horizontal links or ideal（aberration-free） vortex beams, neglecting realistic slant-path conditions and inherent optical aberrations such as coma aberration. This oversight causes a marked decline in performance when applied to real-world systems. Moreover, existing models face difficulty balancing accuracy and efficiency during simultaneous multi-parameter identification. Even in the few studies exploring multi-task learning, the training process often favors the simple task of topological charge classification, neglecting the more challenging task of coma coefficient regression. To address this issue, this paper introduces a dynamic weight learning strategy that adaptively adjusts task weights during neural network training. This approach effectively suppresses the network’s bias towards the dominant task（topological charge recognition）, thereby significantly improving the identification accuracy and convergence speed for the coma coefficient.Methods A slant-path propagation model for vortex beams carrying coma aberrations through atmospheric turbulence was established. Based on the ResNet18 architecture, a dynamic weights strategy was employed for the simultaneous recognition of the topological charge（l） and the coma coefficient（kC₃）. To prevent overfitting and enhance the model’s generalization ability, a Dropout layer was incorporated before the final fully connected linear layer. The topological charge（l） and coma coefficient（kC₃） were both set to integer values from 1 to 10, incremented by 1. The dynamic weight coefficient（α） was initialized at 0.5 and decayed exponentially by 0.95 per epoch until reaching 0.2. Specifically, the weight for the loss associated with l was set to α, while the remaining weight 1-α was allocated to the kC₃ loss. Consequently, α started at 0.5 and progressively decreased each epoch（following 0.5 × 0.95Eepoch）,eventually stabilizing at 0.2. Furthermore, by comparing the dynamic weights strategy with the average weights approach to demonstrate its advantages, and by investigating the impact of different transmission distances and zenith angles on the model’s performance in identifying l and kC₃.Results and Discussions The experimental results demonstrate that the model with dynamic weighting achieved earlier and faster convergence in accuracy. Additionally, the final training duration with dynamic weighting was 746 s, notably shorter than the 844 s required with average weighting. It is noteworthy that both strategies ultimately enabled the model to reach optimal performance on all three evaluation metrics（recall, precision, F1-score） for both l and kC₃ tasks. Additionally, the ResNet18 model employing dynamic weighting achieved perfect scores（1） on all three evaluation metrics for identifying both the topological charge l and kC₃ under a zenith angle of 30° at various transmission distances（2000 m, 3000 m, 4000 m）. As the transmission distance increased, the model’s convergence slowed, and the training process for kC₃ became less stable, exhibiting noticeable oscillations. This indicates that the training difficulty escalates with longer distances. Furthermore, training times for increasing transmission distances were 965 s,1016 s, 1064 s, respectively. To further investigate the impact of the zenith angle, the transmission distance was fixed at 2000 m. At zenith angles of 30°, 45°, 60°, the model maintained perfect recognition metrics（1） for both l and kC₃. However, as the zenith angle increased, the beam experienced more severe turbulent disturbances, resulting in progressively longer training times of 965 s, 1037 s,1074 s, respectively. A comparative analysis of ResNet18 with ResNet34 and VGG16 revealed that ResNet18 achieved superior performance on the three evaluation metrics compared to ResNet34, while requiring a shorter training time than VGG16. Therefore,the ResNet18 network proves to be the most efficient and reliable architecture for this dual-parameter identification task, maintaining high accuracy while achieving faster training times.Conclusions To address the challenge of high-precision dual identification of topological charge and coma coefficient in vortex beams under slant-path atmospheric turbulence transmission, a slant-path transmission model incorporating random atmospheric turbulence perturbations and coma aberration was constructed, generating a dataset comprising nine different parameter combinations. Building upon this, a dynamic weights multi-task learning approach was proposed. This method adaptively adjusts the loss weights of the classification and regression tasks, effectively mitigating the imbalance issue in multi-task learning, and is successfully applied to a ResNet18 network. Experimental results demonstrate that without compromising the speed and accuracy of topological charge identification, the network model employing dynamic weights exhibits a significant improvement in the identification speed of the coma coefficient. The training process is more stable, and the overall training time is reduced. Under various transmission distances and zenith angles, the model maintains high identification accuracy for both parameters despite the influence of these conditions.Comparative experiments under identical conditions with ResNet34 and VGG16 show that ResNet18 retains a significant advantage in this dual-task identification. Although the simultaneous identification of topological charge and coma aberration has been investigated,the actual atmospheric environment is far more complex. It is necessary to consider more realistic atmospheric conditions and to further explore more intelligent and efficient weight adjustment mechanisms.更多还原