摘要
为了快速精准检测无砟轨道伤损,基于无砟轨道表观伤损相机激光一体化成像检测技术、多相机自适应控制和多图像自动拼接技术,研发了无砟轨道表观伤损检测装置,实现了无砟轨道轨道板及层间离缝全断面高清图像的快速采集;提出了多模协同表观图像伤损混合缺陷智能识别技术,在编码-解码语义分割核心结构下引入Transformer模块,并采用TensorRT模型轻量化部署技术实现GPU加速推理,降低计算复杂度,提升并行能力。样本库测试结果表明,本文算法对5 000个测试集样本的识别准确率为95.84%。现场试验结果表明,该装置对无砟轨道表观伤损的检出率为96.4%。无砟轨道表观伤损智能检测装置能够进行无砟轨道表观伤损的高效准确检测,检测方法稳定性、泛化性好。
In order to quickly detect and identify damage of ballastless tracks,an intelligent detection device for apparent damage of ballastless tracks was developed based on the laser integrated imaging detection technology of the apparent damage of ballastless tracks,multi camera adaptive control,and multi image automatic splicing technology.This has achieved rapid collection of high-definition images of the entire cross-section of the ballastless track slab and interlayer gap.A multimodal collaborative apparent image damage mixed defect intelligent recognition technology was proposed.The Transformer module was introduced under the core structure of encoding-decoding semantic segmentation,and the TensorRT model lightweight deployment technology was used to accelerate GPU inference,reduce computational complexity,and improve parallel ability.The sample library test results show that the recognition accuracy of our algorithm for 5000 test samples is 95.84%.The on-site test results show that the detection rate of apparent damage on ballastless track by this device is 96.4%.The intelligent detection device for apparent damage of ballastless tracks can efficiently and accurately detect the apparent damage of ballastless track,and the detection method has good stability and generalization.
作者
王宁
李健超
王智超
柴雪松
暴学志
刘艳芬
WANG Ning;LI Jianchao;WANG Zhichao;CHAI Xuesong;BAO Xuezhi;LIU Yanfen(Railway Engineering Research Institute,CARS,Beijing 100081,China;State Key Laboratory for Track System of High-speed Railway,CARS,Beijing 100081,China)
出处
《铁道建筑》
北大核心
2023年第10期11-15,共5页
Railway Engineering
基金
国家重点研发计划(2022YFB2603302)
中国铁道科学研究院集团有限公司基金(2022YJ330)。
关键词
无砟轨道
表观伤损
精准检测
深度学习
智能识别
轻量化部署
ballastless track
apparent damage
accurate detection
deep learning
intelligent recognition
lightweight deployment
