基于轻量化多尺度神经网络的ZPW-2000移频信号检测方法  

Detection of ZPW-2000 Frequency Shift Signal Based on Lightweight Multi-Scale Neural Network

作　　者：武晓春[1] 刘欣然 WU Xiaochun;LIU Xinran(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou Gansu 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《中国铁道科学》2024年第5期187-197,共11页China Railway Science

基　　金：中国国家铁路集团有限公司科技研究开发计划课题(N2022G012)。

摘　　要：针对ZPW-2000移频信号在不平衡牵引电流干扰时低频信号难以检测的问题,提出基于卷积注意力模块的轻量化多尺度神经网络的移频信号低频信息检测方法。首先,根据ZPW-2000移频信号的载频范围,使用不同卷积核大小的多尺度层提取相应载频调制下的移频信号特征;其次,建立线性倒残差模块实现网络轻量化,在保证网络检测准确率的同时减少网络参数,缩短网络检测时长;最后,引入卷积注意力模块,标定通道和空间特征权重,提升网络性能,通过全连接层进行分类,输出18种低频信号的概率分布。结果表明:将含有工频谐波干扰等5类噪声的移频信号输入低频检测模型中进行检测,平均准确率可达99.22%,召回率达到99.21%,综合评价指标值为0.992,检测时间不超过0.249 s。该方法检测效果更优,具有良好的抗干扰能力,可为带内噪声干扰条件下检测ZPW-2000移频信号的低频信息提供重要参考。To address the challenge in detecting low-frequency signals in ZPW-2000 frequency shift signals under unbalanced traction current interference,lightweight multi-scale neural network approach based on a convolutional attention mechanism is proposed for low-frequency signal detection.First,multi-scale layers with different convolutional kernel sizes are employed to extract frequency-shift signal features across various carrier frequency modulations,based on the carrier frequency range of ZPW-2000 signals.Second,an inverse residual linear bottleneck is introduced to optimize the network’s efficiency,reducing parameters and shortening detection time while maintaining high detection accuracy.Finally,the Convolutional Block Attention Module is introduced to calibrate channel and spatial feature weights,thereby enhancing network performance.Classification is performed through a fully connected layer,producing the probability distribution of 18 types of low-frequency signals.Experimental results indicate that when frequency shift signals containing five types of noise,including power-frequency harmonic interference,are input into the low-frequency detection network,the average accuracy reaches 99.22%,the recall rate 99.21%,and the F1 score 0.992,with detection time not exceeding 0.249 seconds.Compared to traditional detection methods,the proposed method excels in detection performance and robust anti-interference capabilities,providing a valuable reference for detecting low-frequency information of ZPW-2000 frequency-shift signals under in-band noise interference conditions.

关 键 词：轻量化卷积神经网络 谐波干扰 多尺度神经网络 信号检测 ZPW-2000移频信号 

分 类 号：U284.28[交通运输工程—交通信息工程及控制]