Deep learning-based automatic modulation recognition (DL-AMR) methods are mainly based on centralized learning and decentralized learning. These methods have been developed for many applications in heterogeneous wireless networks (HWNs). However, the centralized DL-AMR methods involve high communications cost and computation cost and the decentralized DL-AMR methods are not robust since these methods are based on the assumption of uniform data distribution. In the practical HWNs, most sub-networks are independent and their heterogeneous datasets often do not match. In order to solve these problems, we propose a generalized AMR (GAMR) method based on distributed learning by considering the data mismatch scenario. First, each sub-network trains its own model by means of initialization model download from fusion center and formative modulated datasets. Second, sub-networks upload model weight of all sub-networks to the fusion center for re-training a generalized model, which will be immediately distributed to sub-networks for updating local model. Repeat these two steps until the global model converges successfully. Finally, simulation results are given to confirm advantages of the proposed GAMR method in different scenarios of HWNs.

基于深度学习的自动调制识别（DL-AMR）方法主要基于集中学习和分散学习。这些方法已被开发用于异构无线网络 (HWN) 中的许多应用。然而，集中式 DL-AMR 方法涉及高通信成本和计算成本，而分散式 DL-AMR 方法并不稳健，因为这些方法基于均匀数据分布的假设。在实际的 HWN 中，大多数子网络都是独立的，并且它们的异构数据集通常不匹配。为了解决这些问题，我们考虑了数据不匹配的场景，提出了一种基于分布式学习的广义AMR（GAMR）方法。第一的，每个子网络通过从融合中心和形成性调制数据集下载初始化模型来训练自己的模型。其次，子网络将所有子网络的模型权重上传到融合中心重新训练一个广义模型，该模型将立即分发到子网络以更新本地模型。重复这两个步骤，直到全局模型成功收敛。最后，给出了仿真结果以确认所提出的 GAMR 方法在不同 HWN 场景中的优势。