Automatic modulation classification (AMC) is an essential technology for the non-cooperative communication systems, and it is widely applied into various communications scenarios. In the recent years, deep learning (DL) has been introduced into AMC due to its outstanding identification performance. However, it is almost impossible to implement previously proposed DL-based AMC algorithms without large number of labeled samples, while there are generally few labeled sample and large unlabel samples in the realistic communication scenarios. In this paper, we propose a transfer learning (TL)-based semi-supervised AMC (TL-AMC) in a zero-forcing aided multiple-input and multiple-output (ZF-MIMO) system. TL-AMC has a novel deep reconstruction and classification network (DRCN) structure that consists of convolutional auto-encoder (CAE) and convolutional neural network (CNN). Unlabeled samples flow from CAE for modulation signal reconstruction, while labeled samples are fed into CNN for AMC. Knowledge is transferred from the encoder layer of CAE to the feature layer of CNN by sharing their weights, in order to avoid the ineffective feature extraction of CNN under the limited labeled samples. Simulation results demonstrated the effectiveness of TL-AMC. In detail, TL-AMC performs better than CNN-based AMC under the limited samples. What's more, when compared with CNN-based AMC trained on massive labeled samples, TL-AMC also achieved the similar classification accuracy at the relative high SNR regime.

中文翻译：

---

ZF-MIMO 系统中半监督自动调制分类的迁移学习


自动调制分类（AMC）是非合作通信系统的一项关键技术，广泛应用于各种通信场景。近年来，深度学习（DL）因其出色的识别性能而被引入AMC。然而，在没有大量标记样本的情况下，几乎不可能实现先前提出的基于深度学习的AMC算法，而在实际通信场景中，通常标记样本较少，无标签样本较多。在本文中，我们提出了一种在迫零辅助多输入多输出（ZF-MIMO）系统中基于迁移学习（TL）的半监督AMC（TL-AMC）。 TL-AMC具有新颖的深度重建和分类网络（DRCN）结构，由卷积自动编码器（CAE）和卷积神经网络（CNN）组成。未标记的样本从 CAE 流出以进行调制信号重建，而标记的样本则输入 CNN 以进行 AMC。通过共享权值将知识从CAE的编码器层转移到CNN的特征层，以避免CNN在有限的标记样本下无法有效地提取特征。仿真结果证明了TL-AMC的有效性。具体来说，在有限样本下，TL-AMC 的表现优于基于 CNN 的 AMC。更重要的是，与在大量标记样本上训练的基于 CNN 的 AMC 相比，TL-AMC 在相对高的 SNR 范围内也实现了相似的分类精度。