基于R-P图像注意融合网络的列车轮对踏面缺陷识别  

Wheelset tread defect recognition based on R-P image attention fusion network

作　　者：杨能普 周苗 王文昆 张娟 孙永奎 YANG Nengpu;ZHOU Miao;WANG Wenkun;ZHANG Juan;SUN Yongkui(College of Railway Transportation,Hunan University of Technology,Zhuzhou 412007,China;Zhuzhou CRRC Times Electric Co.,Ltd.,Zhuzhou 412007,China;China Railway Special Transportation of Large Cargo Services Co.,Ltd.,Beijing 100070,China;National Engineering Research Center of Rail Transportation Operation and Control System,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]湖南工业大学轨道交通学院,湖南株洲412007 [2]株洲时代电子技术有限公司,湖南株洲412007 [3]中铁特货大件运输有限公司,北京100070 [4]北京交通大学轨道交通运行控制系统国家工程研究中心,北京100044

出　　处：《铁道科学与工程学报》2023年第12期4811-4822,共12页Journal of Railway Science and Engineering

基　　金：国家重点研发计划资助项目(2021YFF0501101);湖南省自然科学基金资助项目(2021JJ40180)。

摘　　要：轮对踏面缺陷识别是开展列车轮对检修维护的重要前提。然而,轮对踏面缺陷样本存在的数量少、类别不均衡现象,使得传统深度学习方法识别精度有限,难以满足列车智能运维发展的需求。提出一种基于R-P图像注意融合网络(RP-AFN)的列车轮对踏面缺陷识别方法。首先,将RGB图像及其泊松编码(Poisson coding,POS)形式同时引入Mobilenetv2网络之中,以通过丰富输入信息的方式来缓解小样本问题;其次,将挤压-激励模块(SEBlock)和相似性约束引入模型中,以提取具有模态间交互性信息的判别注意力特征,并利用多层感知机(MLP)对其进行融合和推断;最后,基于焦点损失和相似性损失设计一种联合损失函数,使其能够通过惩罚大类别样本以缓解类别不平衡问题。实验结果表明:所提融合方法实现了85.7%的轮对踏面缺陷识别率,且其各项指标以约2.5%~20%的优势优于对比方法。在不同输入大小和主干网络的条件下,以224×224为输入大小的Mobilenetv2模型所对应的模型性能最佳。在不同特征融合策略的条件下,特征拼接方法的各项指标以约1%~3.5%的优势优于特征加法、特征外积和动态加权等方法。在不同调剂因子μ的设置中,μ=0.7的情况下,模型各项指标性能达到综合最优。在主干网络相同的条件下,融合方法能够在不显著增加模型参数的情况下,其各项指标优于单模态方法约1%~6.3%。在执行消融实验的情况下,输入模态、损失函数和模块的消融都会给模型带来负面影响,验证了模型设计的合理性。RP-AFN模型能够有效地提高轮对踏面缺陷识别性能,可在一定程度上解决小样本和类别不均衡条件下的踏面缺陷识别问题。Recognition of wheelset tread defects is an important prerequisite for carrying out train wheelset inspection and maintenance.However,due to the small number and unbalanced categories of wheel tread defect samples,the recognition accuracy of traditional deep learning methods is limited,which makes it difficult to meet the needs of the development of intelligent operation and maintenance of train wheelsets.A R-P image attention fusion network(RP-AFN)-based method was proposed for train wheelset tread defect recognition.First,the RGB images and their Poisson encoding(POS)forms were simultaneously introduced into the Mobilenetv2 network to alleviate the small sample problem by enriching the input information.Second,the squeeze-excitation block(SE Block)and similarity constraints were introduced into the model to extract discriminative attention features with interactive information between modes,which were used for fusion and inference by a multi-layer perceptron(MLP).Finally,a joint loss function was designed based on focal loss and similarity loss,which can alleviate the category imbalance problem by penalizing the large category samples.The experimental results show that the proposed fusion method achieves a wheelset tread defect recognition accuracy of 85.7%.Besides,its various indicators are superior to the comparison methods with advantages of about 2.5%~20%.Under the conditions of different input sizes and backbone networks,the Mobilenetv2 model with an input size of 224×224 performs best.Under the conditions of different feature fusion strategies,the indicators of the feature connection method are superior to the methods of feature addition,feature outer-product and dynamic weighting with advantages of about 1%~3.5%.Under the conditions of different adjustment factorsμ,the performance of various indicators of the model achieves comprehensive optimization whenμis equal to 0.7.Under the conditions of the same backbone network,the fusion method can outperform the single-mode method by about 1%to 6.3%in vario

关 键 词：轮对踏面缺陷识别 图像融合 深度学习 注意力机制 

分 类 号：U279[机械工程—车辆工程]