基于3DCNN-BiConvLSTM的莫尔斯码自动识别算法  

Automatic Recognition Algorithm of Morse Code Based on 3DCNN-BiConvLSTM

作　　者：李子君 魏振华 韩思明 何玉杰 LI Zijun;WEI Zhenhua;HAN Siming;HE Yujie(School of Operational Support,Rocket Force Engineering University,Xi’an 710025,China)

机构地区：[1]火箭军工程大学作战保障学院,陕西西安710025

出　　处：《无线电工程》2023年第8期1862-1868,共7页Radio Engineering

基　　金：国家自然科学基金(62006240)。

摘　　要：莫尔斯电报通常工作在低信噪(SNR)比环境中,其信号易出现码长偏差和频率偏移的问题,现有自动识别算法在上述情况下的识别效果仍有待提升。提出了一种基于深度学习的识别算法,利用小波变换将莫尔斯信号转换为时频图像,通过三维卷积神经网络(3D Convolutional Neural Network,3DCNN)层和双向卷积长短时记忆神经网络(Bidirectional Convolution Long Short-Term Memory Neural Network,BiConvLSTM)捕获莫尔斯码的时空特征并进行预测,借助连接时序分类层(Connectionist Temporal Classification,CTC)实现了端到端的译码。实验结果表明,该算法在更小训练样本和不同信噪比的前提下,对不稳定的莫尔斯信号均能保持98%以上的单词识别准确率,相比现有算法更具鲁棒性。Morse telegraph usually works in a low SNR environment,and its signal is prone to code length deviation and frequency offset.The recognition effect of the existing automatic recognition algorithm in the above cases still needs to be improved.A recognition algorithm based on deep learning is proposed.The Morse signal is transformed into time-frequency image by wavelet transform.The spatial and temporal characteristics of Morse code are captured and predicted by Three-Dimensional Convolution Neural Network(3DCNN)layer and Bidirectional Convolution Long Short-Term Memory Neural Network(BiConvLSTM).Finally,the end-to-end decoding is realized by Connectionist Temporal Classification(CTC)layer.The experimental results show that the algorithm can maintain more than 98%word recognition accuracy for unstable Morse signals under the premise of smaller training samples and different SNR,which is more robust than the existing algorithms.

关 键 词：莫尔斯码 自动识别 深度学习 低信噪比 码长偏差 频率偏移 

分 类 号：TN922[电子电信—通信与信息系统]