基于卷积神经网络与时频分析的水声通信前导信号检测方法

Preamble detection based on CNN and STFT for underwater acoustic communications

前导检测是水声通信中的关键步骤。传统信号检测算法无法有效克服多径和干扰的影响,在复杂水声信道环境下,检测性能会明显下降。卷积神经网络具有强大的表征学习能力,能够通过大量数据自动提取图像特征,在图像识别应用中具有突出优势。以双曲调频信号为研究对象,采用时频分析技术同时获取信号时域和频域信息,得到时频谱图,并通过均值滤波处理提高信噪比。将时频谱图作为卷积神经网络的输入,训练并测试网络,最终得到HFM信号的检测结果。千岛湖试验结果表明,该方法能够大幅提升远距离、强多径条件下前导信号的检测概率,与传统检测算法相比性能优越。

Preamble detection is an important step in underwater acoustic communications.The traditional detection methods cannot effectively overcome the multipath and interference,so the detection performance will be significantly decline when applied in a complex channel.Convolutional neural network(CNN)has outstanding advantages in the image recognition because of its strong learning ability and automatic feature extraction.Take the HFM signal as the research object,use the time-frequency analysis technology to get the information of time and frequency simultaneously,then obtain the time-frequency spectrum.The mean-filtering is carried out to reduce the image noise.Take the spectrum as the input of CNN,train and test the network,then obtain detection result.The experimental results in Thousand island lake show that the detection method based on CNN and STFT can greatly improve the detection probability in long distance and strong multipath channel,and has superior performance compared with the traditional methods.