基于改进SNN的列车轮对踏面缺陷识别方法  

Defect Detection of Train Wheelset Tread Based on Improved Spiking Neural Network

作　　者：何静[1] 黄聪聪 张昌凡[2] 贾林[2] HE Jing;HUANG Congcong;ZHANG Changfan;JIA Lin(College of Electrical and Information Engineering,Hunan University of Technology,Zhuzhou 412000,China;College of Railway Transportation,Hunan University of Technology,Zhuzhou 412000,China)

机构地区：[1]湖南工业大学电气与信息工程学院,湖南株洲412000 [2]湖南工业大学轨道交通学院,湖南株洲412000

出　　处：《铁道学报》2025年第1期91-100,共10页Journal of the China Railway Society

基　　金：国家自然科学基金(52172403,62173137,62303178)。

摘　　要：列车轮对踏面缺陷识别是保障列车轮轨系统安全服役的关键步骤。然而,轮对踏面缺陷类型多样复杂、类别不平衡,现有卷积神经网络算法难以对其进行准确识别。为此,提出基于改进脉冲神经网络(SNN)的列车轮对不平衡踏面缺陷识别方法。采用混合卷积编码模块,通过提高特征多样性稀疏表达,减少编码细节信息丢失;提出脉冲金字塔拆分注意网络,考虑多尺度空间信息跨通道交互能力,以提取缺陷的多尺度特征;提出一种新的交叉注意力模块,提取不同层级特征的空间全局信息,通过交叉校准以增强输入特征,抑制噪声等无用特征;通过不平衡比例达10∶1的踏面缺陷数据集对该识别方法进行试验验证。验证结果表明,该方法能够有效提高模型的识别精度,并且对少数类别缺陷也有较高的识别率。Wheel tread defect detection is a key step for the safe operation of the wheel⁃rail system.However,due to the diversity and complexity of such defects,as well as the imbalanced dataset,existing detection algorithms based on deep convolutional neural networks can hardly achieve accurate detection.To solve this problem,an improved spiking neural network(SNN)based detection method was proposed for wheel tread defects with imbalanced dataset.Firstly,a hybrid convolutional encoding module was adopted to improve feature diversity and sparse representation,reducing the loss of encoding details.Secondly,a pulse pyramid splitting attention network was proposed to extract multi⁃scale features of de⁃fects by considering the cross channel interaction ability of multi⁃scale spatial information.Thirdly,a new type of cross attention module was proposed to extract spatial global information of features at different levels,strengthen input features through cross calibration,and suppress useless features such as noise.Finally,the advantages of the proposed model were verified using the tread defect datasets with an imbalance ratio of 10∶1.The experimental results show that the proposed method can effectively improve the accuracy of defect detection of the model with a high detection accuracy for minority defects.

关 键 词：轮对踏面 缺陷识别 脉冲神经网络 特征融合 注意力机制 

分 类 号：U45[建筑科学—桥梁与隧道工程]