基于稀疏深度神经网络的电磁信号调制识别

为在低复杂度约束条件下提升电磁信号调制识别的性能，提出了一种基于稀疏深度神经网络(Sparse Deep Neural Network, SDNN)的电磁信号调制识别方法。首先，通过提取电磁信号同相和正交两路数据绘制出信号的星座图，作为信号的浅层特征表达；然后，基于星座图中各信号点密度大小对星座图进行上色，增强星座图中信号特征；最后，通过SDNN对增强后的星座图进行识别分类。实验结果表明，SDNN模型选取合适的剪枝率后，能够有效降低模型存储规模和计算量，其中模型参数压缩了72%,浮点运算量压缩了45%,与原模型97%的综合识别率相比，稀疏化处理后模型的综合识别率为96.8%,在小幅度识别精度损失范围内大幅降低了模型复杂度。

Modulation recognition of electromagnetic signal based on sparse depth neural network

In order to improve the performance of electromagnetic signal modulation recognition under low complexity constraints,an electromagnetic signal modulation recognition method based on sparse depth neural network(SDNN) is proposed.First,by extracting the in-phase/quadrature data of the electromagnetic signal,the constellation of the signal is drawn as the shallow feature expression of the signal.Then,the constellation is colored based on the density of each signal point in the constellation to enhance the signal characteristics in the constellation.Finally,the enhanced constellation is recognized and classified by SDNN.The experimental results show that the SDNN model can effectively reduce the storage scale and computation of the model after selecting the appropriate pruning rate.And the model parameters are compressed by 72% and the floating-point operation is compressed by 45%.Compared with the 97% comprehensive recognition rate of the original model,the comprehensive recognition rate of the sparse model is 96.8%,which is within the range of small loss of recognition accuracy.The complexity of the model is greatly reduced.