基于深度学习的接触网顶紧螺栓状态智能检测

Research on Intelligent Detection of State of Catenary Puller Bolt Based on Deep Learning

高速铁路接触网支撑装置顶紧螺栓在列车长期运行中会产生松脱和脱落等不良状态,针对其缺陷样本不足、状态变化特征粒度差别小不易判定等问题,提出SSD512和U-net8两种深度学习算法。并基于SSD512定位算法和U-net8语义分割模型的设计,实现顶紧螺栓状态的智能检测。通过SSD512实现了顶紧螺栓区域的定位和截取;通过U-net8将顶紧螺栓图片中薄螺母、螺杆等语义信息进行不同颜色的标记;通过语义图片的判定,实现对顶紧螺栓的状态检测。通过在顶紧螺栓的定位、顶紧螺栓的语义分割两个数据集上进行训练、验证,结果表明:所提的顶紧螺栓状态智能检测方法能准确定位接触网中的顶紧螺栓;实现顶紧螺栓的状态检测,综合准确率达到95.75%。

The puller bolts of the support device of high-speed railway catenary may be loosened or fall off during the long-term operation of the train.In order to address the difficulty in determining these problems due to insufficient defect samples and the small granularity of state changes in the picture,two deep learning algorithms named SSD512 and U-net8 were proposed.Based on the SSD512 positioning algorithm and the design of the U-net8 semantic segmentation model,the intelligent detection of puller bolt state was realized.Firstly,the target detection algorithm called SSD512 was used to locate the puller bolt area.Then,the semantic segmentation algorithm called U-net8 was used to mark the semantic information such as thin nuts and screws in the puller bolt pictures in different colors.Through the judgment of semantic picture,the state detection of puller bolt was realized.Training and testing were performed on the two datasets,namely the locating of the puller bolts and the semantic segmentation of the puller bolts.The experimental results show that the method proposed can achieve a comprehensive accuracy of 95.75%in the intelligent state detection of catenary puller bolts.