基于全卷积神经网络的机车信号降噪

机车信号从钢轨提取轨道电路信号作为行车凭证，其译码输出性能对列控系统的可靠性和安全性有直接影响. 但列车运行过程中，机车信号不可避免地混入大量噪声和干扰，译码前需要降噪以提高准确性. 为此，提出一种基于全卷积神经网络（fully?convolutionalnetworks,?FCN）的机车信号降噪方法，该方法利用基于原始波形“端到端”处理方式的FCN，直接从时域对机车信号进行降噪处理，以提高信噪比（signal-to-noise?ratio，SNR）；并利用仿真和实测数据对本方法进行了实验. 结果表明:相较于传统基于频谱的滤波方法，本方法对带内干扰有更显著的效果，采用FCN能使机车信号信噪比提高8～14 dB，可有效降低带内噪声. 

Cab Signal Denoising Process Based on Fully Convolutional Networks

Since cab signals extract information from track circuits as the running token, its decoding performance has a direct impact on the reliability and security of train operation control system. However, as it is inevitable that a lot of noise and interference will mix into the cab signal during operation, it is necessary to denoise before decoding in order to improve demodulation accuracy. To this end, a raw waveform-based fully convolutional network (FCN) for denoising is proposed in an end-to-end manner, which denoises the cab signal in time domain directly and improves the signal-to-noise ratio (SNR). This proposed network is validated through simulation and measured data. The experimental results show that compared with the traditional spectrum-based denoising methods, this method has a more significant effect on in-band interference; FCN can improve the SNR of cab signals by 8～14 dB and effectively reduce the in-band interference.