秦菲菲,董昱.面向列车轮轨接触区域图像分割的生成对抗网络*[J].电子测量与仪器学报,2021,35(11):154-162 |
面向列车轮轨接触区域图像分割的生成对抗网络* |
Generative adversarial network for image segmentation of train wheel rail contact area |
|
DOI: |
中文关键词: 轮轨接触区域边缘曲线分割 生成对抗网络 卷积神经网络 对抗学习 |
英文关键词:edge curve segmentation of wheel rail contact area generative adversarial network convolutional neural network adversarial learning |
基金项目:国家自然科学基金(61763023)项目资助 |
|
|
摘要点击次数: 2124 |
全文下载次数: 716 |
中文摘要: |
列车和轨道之间的开放约束条件决定了车辆脱轨的客观存在。轮轨接触区域边缘曲线分割对列车轮轨接触关系的研究具有重要意义,提出了一种基于生成对抗网络的轮轨接触区域边缘曲线分割算法。通过将残差模块引入生成器网络中,增强了网络对输出变化的敏感程度,进而更好的调整生成器权重。此外,膨胀残差模块的引入,有效扩大了特征图的接收区域。实验结果显示,改进的生成对抗网络对轮轨接触区域边缘曲线的分割准确度达到9613%,敏感度、特异度、F1值、ROC曲线下的面积分别为8390%、9713%、8367%和9812%,验证了该方法能够准确分割轮轨接触区域边缘曲线。 |
英文摘要: |
The open constraint condition between the train and the track determines the objective existence of the train derailment. Curve segmentation of the edge of wheel rail contact area is of great significance to the research of the train wheel rail contact relationship. In this paper, an algorithm for the curve segmentation of the edge of wheel rail contact area based on generative adversarial networks is proposed. By introducing the residual module into the generator network, the sensitivity of the network to output changes is enhanced, and the generator weight can be better adjusted. In addition, in order to effectively expand the receiving area of the feature map, the expansion residual module is introduced. The experimental results show that the accuracy of curve segmentation of the edge of wheel rail contact area reaches 9613% by improved generative adversarial networks, and the sensitivity, specificity, F1 value and area under the ROC curve is 8390%, 9713%, 8367% and 9812% respectively, which verify that this method can accurately segment the edge curve of the wheel rail contact area. |
查看全文 查看/发表评论 下载PDF阅读器 |