基于网络蒸馏模型的供电接触网支撑装置零部件快速定位方法
喻文彬
Quick Positioning Method of Catenary Support Device Components Based on Network Distillation Model
YU Wenbin
-
作者信息:中铁建电气化局集团第四工程有限公司，410116，长沙
-
Affiliation:The 4th Engineering Co., Ltd. of China Railway Construciton Electrification Bureau Group Co., Ltd., 410116, Changsha, China
-
关键词:
-
Key words:
-
DOI:10.16037/j.1007-869x.2023.05.021
-
中图分类号/CLCN:U225.4
-
栏目/Col:研究报告
摘要:
为确保供电接触网支撑装置具备良好的工作状态，避免支撑装置失效可能造成的严重弓网故障，进而影响列车运行安全，提出一种新型的、适用于小型化嵌入式设备进行部署应用的、基于网络蒸馏技术的深度学习接触网支撑装置快速定位识别模型。网络蒸馏模型是基于双深度学习网络进行指导性学习的，其利用高精度深度学习模型YOLOv5x、YOLOv5l、YOLOv5m作为蒸馏算法中的教师模型，然后利用网络蒸馏技术对复杂模型的特征提取能力进行迁移，以教师模型的类别概率作为软目标并将其用于训练学生模型，从而将知识从复杂模型（教师模型）转移到更高效的小型化模型（学生模型）。研究结果表明，利用网络蒸馏技术可以获得精度高、速度快、模型小、易于部署的接触网支撑装置定位模型。
Abstracts:
To ensure that catenary support device is in good working condition and avoid serious pantograph-catenary faults caused by support device failure, which may further undermine railway operation safety, a new deep learning catenary support device quick positioning and recognition model is proposed, that is based on network distillation technology and can be deployed on embedded devices. The network distillation model achieves guided learning through dual deep learning networks. It takes the high-precision deep learning models YOLOv5x, YOLOv5l and YOLOv5m as teacher models in the distillation algorithm, and then the network distillation technology is used to migrate the feature extraction capability of complex model, the category probabilities of teacher model are used as soft targets for the training student model, so that the knowledge is transferred from the complex model (teacher model) to an efficient miniaturized model (student model). Research results show that the network distillation technology can obtain an accurate, fast, small, and easy-to-deploy model for positioning catenary support device.
过刊浏览
资料下载
友情链接