基于卷积自编码网络的BPSK信号降噪

BPSK Signal Denoise Based on Convolution Auto-Encoder Network

针对传统降噪方法提取特征繁琐、参数选取不易的问题,提出了基于卷积自编码器(Convolutional Autoencoder,CAE)降噪的方法,对BPSK信号、加性高斯白噪声,信噪比-10dB 2dB的数据构成信号数据集。在网络训练阶段,将加噪后的样本经过卷积自编码器提取潜在特征,多次训练迭代并且保存模型的参数;在测试阶段,利用新产生的测试集完成对该算法的验证与测试,可以观察到恢复出的有用信号,且误码率有了明显的降低。实验表明,相对于传统信号降噪算法(例如小波阈值降噪、PCA等),所提算法不需要人工手动提取信号特征,实现了对BPSK信号的降噪处理。

Aiming at the problem of tedious feature extraction and difficult parameter selection of traditional noise reduction methods, a method based on Convolutional Autoencoder(CAE) noise reduction is proposed. For BPSK signal, additive Gaussian white noise, signal-to-noise ratio-10 dB-2 dB data constitutes signal data set. In the network training phase, the noisy samples are extracted by the convolutional self-encoder to extract potential features, and the training parameters are iterated multiple times and the parameters of the model are saved;in the test phase, the newly generated test set is used to complete the verification and test of the algorithm The useful signals recovered can be observed, and the bit error rate has been significantly reduced. Experiments show that compared with traditional signal noise reduction algorithms(such as wavelet threshold noise reduction, PCA,etc.), the proposed algorithm does not need to manually extract signal characteristics manually, and achieves noise reduction processing for BPSK signals.