基于相位融合特征和残差网络的调制智能识别

Modulation intelligent recognition based on phase fusion feature and residual network

针对调制识别依赖人工提取的经验特征、识别准确率低的问题,提出基于相位星座图和矢量轨迹图融合聚类的智能识别算法.通过对接收信号进行滤波、定时同步预处理,恢复出较理想的基带信号,将信号的波形数据转化为星座图和矢量轨迹图特征.利用深度学习(DL)将调制识别转换成图像的分类问题,将所提取的特征通过2路并行输入的轻量级残差结构网络,开展分层学习和特征融合训练,完成对目标调制方式的识别.仿真实验表明,基于融合特征的识别结果优于目前的基于高阶累积量、星座图和波形数据的识别结果,当信噪比(SNR)高于2 dB时,对MPSK(调制的阶数为2、4、8)、MQAM(调制的阶数为16、64)、MAPSK(调制的阶数为16、32)这7类调制的识别率可以达到95.14%.

An intelligent signal modulation recognition algorithm based on phase constellation and trajectory clustering was proposed aiming at the problem that the common recognition relied on manually extracted empirical features and suffered from low precision. An ideal baseband signal was recovered by preprocessing the received signal, such as filtering and timing synchronization. Then the waveform data of the signal was transformed into a constellation diagram and vector trajectory diagram. The signal modulation recognition problem was transformed into a featured image classification problem by exploiting the techniques in deep learning(DL). The extracted features were input into a lightweight residual structure network with two parallel inputs. Then hierarchical learning and feature fusion training were performed to achieve the target modulation recognition. The simulation experiments showed that the recognition rates achieved by the feature fusion method outperformed those achieved by the methods based on high-order statistics, constellation diagrams and signal waveform. The recognition rate for seven types of modulation consisting of MPSK(M = 2, 4, 8), MQAM(M = 16, 64), and MAPSK(M = 16, 32) can reach 95.14%when the signal-to-noise ratio(SNR) is larger than 2 dB.