恶劣成像环境下高铁弓网图像深度学习与燃弧识别

The pantograph-catenary system is responsible for power transmission to the CHR (China High-speed Railway). With the rapid increase of high-speed railway operating mileage, how to perceive pantograph state and warn fault accurately and efficiently becomes a huge challenge facing the development of high-speed rail. Current pantograph-catenary state monitoring only uses visible image, and cannot obtain the clear image and temperature field distribution of the pantograph-catenary in the poor imaging environment. Therefore, it is urgent to solve the problem of pantograph-catenary state monitoring through heterogeneous image fusion. This project proposes a multi-model fusion method based on deep learning, and studies the image deep learning and arcing recognition from three aspects , namely the basic theory, key technologies and practical examples. First of all, to study the theory of multi-model fusion, and propose the image feature extraction, association learning and fusion inference model based on deep learning. Second, to study the following key technologies for online perception of pantograph-catenary state: arcing recognition based on the multi-instance and multi-label classification model, image registration based on deep feature and heterogeneous image fusion based on multi-mode learning. Finally, combined with high speed railway pantograph-catenary monitoring tasks based on heterogeneous images, to verify the proposed theory and key technology. Through the above research, to solve high-speed railway pantograph-catenary online monitoring problem under poor imaging environment, has great significance to the theoretical research and application of the deep learning.

高速铁路弓网系统担负将电能输送给动车组的重要任务，随着高速铁路运营里程的快速增长，如何对大量动车组在途弓网状态进行高效准确地感知与故障预警成为高铁发展面临的挑战。现有弓网状态监测仅利用可见光单模态视频图像，无法获取恶劣成像环境下弓网清晰图像与温度场分布，因此，亟需解决异质图像融合的弓网状态监测难题。本项目旨在提出一种基于深度学习的异质图像融合方法，从基础理论、关键技术和实例验证三个方面研究弓网图像深度学习和燃弧识别问题。首先，研究异质图像多模融合理论，提出基于深度学习的图像特征提取、关联学习和融合推理模型；其次，研究多实例多标签的燃弧类别识别、深度特征的图像配准和多模学习的异质图像融合的关键技术，实现弓网状态的在线感知；最后，结合承担的异质图像高铁弓网监测任务，验证提出的理论与关键技术。通过上述研究，探索解决恶劣成像环境下高铁弓网状态在线监测难题，对深度学习理论研究和应用具有重要意义。

高速铁路弓网系统担负将电能输送给动车组的重要任务，随着高速铁路运营里程的快速增长，如何对大量动车组在途弓网状态进行高效准确地感知与故障预警成为高铁发展面临的挑战。现有弓网状态监测仅利用可见光单模态视频图像，无法获取恶劣成像环境下弓网清晰图像与温度场分布，因此，亟需解决异质图像融合的弓网状态监测难题。本课题实现了高速列车接触网的燃弧识别、红外图像配准和运行在线监测的三个目标。首先，在理论方法创新方面，1）构建了高铁接触网燃弧数据集，并提出了基于燃弧形态的燃弧分类方法，2）提出了基于深度卷积特征的可见光与红外图像配准方法。其次，在关键技术突破方面，1）提出了基于单目视觉的高铁接触网定位器坡度计算方法，2）提出了基于改进RetinaNet的高铁接触网鸟巢检测方法，3）提出了基于“全局检测+局部识别”的高铁接触网杆号识别方法，用于对接触网燃弧进行精准定位。最后，在验证性应用方面，本课题的关键技术均编写了算法模块,并应用于成都国铁电器公司的高铁接触网在线检测设备上，成功服务于国家高速列车接触网的安全监测，并且在实际应用中获得了极高的评价。