非均衡数据下基于注意力网络和代价敏感学习的轨面状态识别  

Rail Surface State Recognition Based on Attention Network and Cost-sensitive Learning under Imbalanced Data

作　　者：于惠钧[1] 张锦圣 刘建华[1] 彭慈兵 刘丽丽[2] 龚事引 YU Hui-jun;ZHANG Jin-sheng;LIU Jian-hua;PENG Ci-bing;LIU Li-li;GONG Shi-yin(College of Railway Transportation,Hunan University of Technology,Zhuzhou 412007,China;College of Intelligent Control,Hunan Railway Professional Technology College,Zhuzhou 412012,China;College of Railway Transportation and Electrical Engineering,Hunan Vocational College of Railway Technology,Zhuzhou 412000,China)

机构地区：[1]湖南工业大学轨道交通学院,株洲412007 [2]湖南铁道职业技术学院智能控制学院,株洲412012 [3]湖南铁路科技职业技术学院铁道供电与电气学院,株洲412000

出　　处：《科学技术与工程》2024年第5期1972-1979,共8页Science Technology and Engineering

基　　金：国家重点研发项目(2021YFF0501101);国家自然科学基金(52272347);湖南省教育厅科学研究重点项目(20A162);湖南省自然科学基金(2021JJ30217,2022JJ50095)。

摘　　要：准确识别轨面状态,可为列车牵引/制动性能提升提供关键依据。重点针对传统代价敏感学习应用在非均衡轨面状态识别中存在的同类别样本重要性不同和多数类精度下降等问题,提出一种基于注意力网络和代价敏感学习的轨面状态识别方法。该法首先利用迁移学习思想将均衡数据集的特征迁移到非均衡轨面状态数据集,减轻少数类样本误分类影响;其次在骨干网络ResNet18中引入卷积注意力机制模块,增强网络对目标区域的特征学习能力和全局特征信息的感知性能,调整优化网络权重参数;最后构造依据轨面状态样本重要性大小的自适应加权平衡损失函数,降低决策边界对困难样本中多数类的过拟合,获得更加平滑的决策边界。非均衡数据下的实验结果表明,在3种非均衡比下,所提方法的准确率和召回率分别达到96.00%、90.67%、86.33%,与目前常用的方法Focal相比,分别提升了7.00%、2.34%、3.00%。此外,该方法在提高少数类召回率的同时可有效维持多数类的召回率,并且降低了网络训练时间成本。Accurately identifying rail surface state can provide key evidence for improving the traction/braking performance of trains.Focusing on the problems of differing importance of samples within the same class and decreased accuracy of the majority class when traditional cost-sensitive learning is applied to imbalanced rail surface state recognition,a rail surface state classification method based on attention networks and cost-sensitive learning was proposed.Firstly,transfer learning was utilized to transfer features from a balanced dataset to an imbalanced rail surface state dataset,alleviating the impact of misclassification in the minority class.Secondly,a convolutional block attention module was introduced into the ResNet18 backbone network to enhance the feature learning capability within the target region and the perceptual ability of global feature information,while adjusting and optimizing the network s weight parameters.Finally,an adaptive weighted balanced loss function was constructed based on the importance of rail surface state samples,reducing the overfitting of the decision boundary to the majority class in hard samples and obtaining a smoother decision boundary.Experimental results on imbalanced data demonstrate that the proposed method achieves accuracy and recall of 96%,90.67%,and 86.33%respectively under three different imbalanced ratios.Compared to commonly used methods Focal,the proposed method exhibits improvements of 7%,2.34%,and 3%in terms of both accuracy and recall respectively.Furthermore,the method effectively maintains the recall of the majority class while improving the recall of the minority class,and reduces the training time cost of the network.

关 键 词：轨面状态识别 非均衡数据 代价敏感学习 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]