数字信号调制识别下坐标注意力机制方案研究

A Study of Coordinate Attention Mechanism Scheme Under Digital Signal Modulation Recognition

针对低信噪比下神经网络难以提取数字信号空间特征的问题,提出一种基于坐标注意力机制的数字信号识别方案。将8种数字信号进行正交调制,根据其幅度、相位信息序列进行预编码处理,在不同的训练步长下,提取分析数字信号幅度和相位的关键特征,选取合适的神经网络超参数,使网络达到拟合面。坐标注意力机制将数字信号特征进行2个一维特征编码,分别沿纵向和横向捕获幅度和相位的远程依赖关系;将生成的数字信号特征编码为一对方向感知和位置敏感的权重系数,进行数字信号特征的重标定。仿真结果表明,8种数字信号下,调制方式识别率高于95%时,卷积神经网络(Convolutional Neural Network,CNN)中坐标注意力机制信噪比增益约为4 dB,残差神经网络中坐标注意力机制信噪比增益约为8 dB。坐标注意力机制取得了较高的识别率以及更好的信噪比增益,与通道注意力机制、空间注意力机制相比更适用于数字信号解调的应用。

To solve the problem that it is difficult for neural network to extract spatial features of digital signals under low signal-to-noise ratio,a digital signal recognition scheme based on coordinate attention mechanism is proposed.Eight kinds of digital signals are orthogonally modulated,pre-encoded according to their amplitude and phase information sequences,and the key features of amplitude and phase of the digital signals are extracted and analyzed under different training steps,and the appropriate hyperparameters of the neural network are selected to make the network to reach the fitting surface.The coordinate attention mechanism encodes the digital signal features into two one-dimensional features to capture the remote dependence of amplitude and phase along the longitudinal and transverse directions,respectively;and the generated digital signal features are encoded into a pair of direction-aware and position-sensitive weight coefficients for the recalibration of the digital signal features.Simulation results show that:under eight digital signals,when the recognition rate of the modulation method is higher than 95%,the signal-to-noise gain of the coordinate attention mechanism in the Convolutional Neural Network(CNN)is about 4 dB,and the signal-to-noise gain of the coordinate attention mechanism in the residual neural network is about 8 dB.The coordinate attention mechanism achieves a higher recognition rate as well as a better signal-to-noise gain,and is more suitable for digital signal demodulation applications,compared with the channel attention mechanism and the spatial attention mechanism.