基于自注意力的双流结构门控循环单元的自动调制识别

Automatic Modulation Recognition Using Dual-Steam Gated Recurrent Unit Based on Self-Attention

摘要: 自动调制识别是各种无线通信场景中的一项重要任务。受通信环境日益复杂的影响，如何有效、高效地提取无线电信号的时间特征是当前亟待解决的问题。为了解决这个问题，本文提出了一种双流结构的基于自注意力的门控循环单元(Gated Recurrent Unit, GRU)模型。本文的方法首先使用卷积层提取浅层空间特征，然后将特征分为两部分，在两个并行流中使用堆叠的GRU来充分提取时间特征。自注意力的加入使网络能够关注输入序列的不同部分，从而提高特征提取的能力。本文提出的模型在基准数据集上对信噪比(Signal-to-Noise Ratio, SNR)大于等于−6 dB的信号实现了最高的识别准确率。特别地，当SNR为−6 dB时，识别准确率达到了58.9%，比其他的模型提高了5%以上。此外，本文提出的模型具有很强的适应性，即使在较小的数据集上训练也能达到最高的准确率。本文还研究了初始学习率对模型性能的影响，给出了模型达到最高准确率和最高效率的初始学习率。

Abstract: Automatic modulation recognition is an important task in various wireless communication scenari-os. Due to the increasingly complex communication environment, how to extract the temporal characteristics of radio signals effectively and efficiently is an urgent problem to be solved. To solve this problem, this paper proposes a dual-stream gated recurrent units (GRUs) model based on self-attention. The approach first uses the convolutional layers to extract the shallow spatial fea-tures, then splits the features into two parts and uses stacked GRUs in two parallel streams to fully extract the temporal features. The addition of self-attention enables the network to focus on differ-ent parts of the input sequences, thus improving the ability of feature extraction. Our model achieves the highest recognition accuracy for signals with a signal-to-noise ratio (SNR) greater than or equal to −6 dB on the baseline dataset. In particular, when the SNR is −6 dB, the recognition ac-curacy reaches 58.9%, which is more than 5% higher than other models. In addition, our model has strong adaptability and can achieve the highest accuracy even when trained on a small dataset. The paper also studies the impact of the initial learning rate on the model performance and gives the initial learning rate for the model to achieve the highest accuracy and efficiency respectively.