基于改进DnCNN的机车信号抗干扰算法

Anti-interference algorithm for cab signalling based on improved DnCNN

在高速和重载铁路牵引电流干扰严重区段,机车信号对于谐波干扰难以采用传统的时域或频域滤波方法有效抑制,导致掉码、延迟上码、误码等现象出现,影响运营效率.为此,本文首先改进去噪卷积神经网络(Denoising Convolutional Neural Networks,DnCNN),将网络中堆叠的单一尺度的卷积核替换为多尺度卷积核,在保证网络性能的同时降低网络深度,并通过残差学习方式得到预估的噪声分布,进而通过对消的方式抑制落入有效频带中的骚扰,使FSK(Frequency-shift Keying)信号低频幅值在有用频带附近最大化,提高信噪比(Signal-to-Noise Ratio,SNR).通过将本方法与常见的信号去噪算法进行仿真对比分析,并利用现场实测信号解码验证,结果表明,该方法能提升信噪比约13 dB,可以更准确地提取FSK机车信号特征频率,为提高机车信号抗干扰性能提供了新的途径.

In high-speed and heavy-haul railway sections where traction current is seriously interfered, it is difficult to effectively suppress harmonic interference in cab signals by traditional time domain or frequency domain filters, resulting in failures like code missing, code delaying or error coding, which affect the operation efficiency. This paper firstly improves the Denoising Convolutional Neural Network(DnCNN) by using multi-scale convolution kernels instead of single-scale ones, consequently the new algorithm can reduce the network depth while ensuring the denoising performance. Then the estimated noise distribution is obtained by residual error learning so that the in-band disturbance can be suppressed by cancellation, and the Frequency-shift Keying signal near modulated frequency is maximized, therefore raising the Signal-to-Noise Ratio(SNR). By simulating and comparing this method with common signal denoising algorithms, and verifying it by decoding signals measured in the field, results show that this method improves the SNR up to 13 dB, and extract the characteristics of frequency shift signal more accurately, providing a novel method to mitigate the interference in the cab signal.