Automatic modulation classification (AMC) is a critical step to recognize the unknown signal modulation types. It is widely applied in non-cooperative communication systems with diverse modulation types and complex communication environments. In recent years, existing deep learning based AMC methods have been developed to improve the performance of classification in additive white Gaussian noise (AWGN) or Rayleigh channel. However, these methods do not consider the interference problem caused by the Doppler shift in high mobile communication environment. In this paper, we propose a deep learning based robust AMC (RAMC) method to effectively suppress the Doppler shift and achieve classification performance. The proposed convolutional and recurrent fusion network (CRFN) consists of the combination of convolutional neural network (CNN) and simple recurrent unit (SRU) to classify eight types of modulation signal under additive white Gaussian noise (AWGN) and different Doppler shifts. We consider four kinds of fusion networks, i.e., SRU series CNN (SSC), CNN series SRU (CSS), CNN parallel SRU (CPS) and weight average (WA). Simulation results show that the proposed CRFN-CSS method achieves the best performance and brings better performance than either CNN or SRU in 100 Hz Doppler shift. What is more, CRFN-CSS is also robust in the small range of Doppler shifts.

自动调制分类（AMC）是识别未知信号调制类型的关键步骤。它广泛应用于具有多种调制类型和复杂通信环境的非合作通信系统中。近年来，已经开发了基于深度学习的现有AMC方法，以改善加性高斯白噪声（AWGN）或瑞利信道中的分类性能。但是，这些方法没有考虑在高移动通信环境中由多普勒频移引起的干扰问题。在本文中，我们提出了一种基于深度学习的鲁棒AMC（RAMC）方法，以有效地抑制多普勒频移并实现分类性能。所提出的卷积和递归融合网络（CRFN）由卷积神经网络（CNN）和简单递归单元（SRU）组成，用于在加性高斯白噪声（AWGN）和不同的多普勒频移下对八种类型的调制信号进行分类。我们考虑了四种融合网络，即SRU系列CNN（SSC），CNN系列SRU（CSS），CNN并行SRU（CPS）和加权平均（WA）。仿真结果表明，所提出的CRFN-CSS方法在100 Hz多普勒频移方面达到最佳性能，并且比CNN或SRU都具有更好的性能。而且，CRFN-CSS在较小的多普勒频移范围内也很稳定。SRU系列CNN（SSC），CNN系列SRU（CSS），CNN并行SRU（CPS）和权重平均值（WA）。仿真结果表明，所提出的CRFN-CSS方法在100 Hz多普勒频移方面达到最佳性能，并且比CNN或SRU都具有更好的性能。而且，CRFN-CSS在较小的多普勒频移范围内也很稳定。SRU系列CNN（SSC），CNN系列SRU（CSS），CNN并行SRU（CPS）和权重平均值（WA）。仿真结果表明，所提出的CRFN-CSS方法在100 Hz多普勒频移方面达到最佳性能，并且比CNN或SRU都具有更好的性能。而且，CRFN-CSS在较小的多普勒频移范围内也很稳定。