基于深度学习的动车组制动盘闸片视觉定位检测
管春玲1许迎杰2宋跃超3
Visual Positioning Detection of EMU Brake Pad Based on Deep Learning
GUAN Chunling1XU Yingjie2SONG Yuechao3
-
作者信息:1.广州铁路职业技术学院车辆学院，510430，广州
2.中国铁路广州局集团有限公司广州动车段，510665，广州
3.北京纵横基机电科技有限公司，100094，北京
-
Affiliation:1.Guangzhou Railway Polytechnic Locomotive & Car College, 510430, Guangzhou, China
2.China Railway Guangzhou Group Co., Ltd., Guangzhou EMU Depot, 510665, Guangzhou, China
3.Beijing Zongheng Electro-Mechanical Technology Co., Ltd., 100094, Beijing, China
-
关键词:
-
Key words:
-
DOI:10.16037/j.1007-869x.2024.12.011
-
中图分类号/CLCN:U279.323
-
栏目/Col:学术专论
摘要:
［目的］随着轨道交通行业的不断发展，其运营规模和安全保障要求不断提高。为了解决动车组制动盘闸片人工检测效率低下、漏检、误检等问题，有必要研究动车组制动盘闸片视觉定位检测方法。［方法］提出一种基于深度学习的动车组制动盘闸片视觉定位检测方法；针对实际场景中图像特征不明显的问题，以FasterR-CNN(快速区域卷积神经网络)算法为基础，引入边缘检测分支，同时在损失函数中添加目标边缘损失函数，合并附加辅助网络的边缘信息。采用双线性插值法计算特征像素值，保留更多动车组制动盘闸片特征信息。［结果及结论］所提改进的FasterR-CNN模型能够在边缘处进行细节处理，加快网络的收敛速度，学习更多动车组制动盘闸片的边缘特征。通过双线性插值法减小了ROI(检测候选框区域)池量化过程中目标特征的错位误差。所提闸片检测方法的平均精度为98.42%，FPS(每秒帧数)为27.77%。
Abstracts:
［Objective］ With the continuous development of the rail transit industry, its operation scale and safety requirements are constantly increasing. To address the issues of low efficiency, missed and false detections in the manual inspection of EMU (electric multiple unit) brake pads, it is necessary to study visual positioning detection methods for EMU brake pads. ［Method］ A visual positioning detection method for EMU brake pads based on deep learning is proposed. To address the issue of indistinct image features in real-world scenarios, based on the Faster R-CNN (region conventional neural network) algorithm, an edge detection branch is introduced, and a target edge loss function is added to the loss function, integrating edge information from an auxiliary network. Bilinear interpolation is used to calculate feature pixel values, preserving more feature information of the EMU brake pads. ［Result & Conclusion］ The improved Faster R-CNN model proposed here can handle details at the edges, accelerate network convergence, and learn more edge features of the EMU brake pads. Through the bilinear interpolation method, the misalignment errors of target features during ROI (region of interest) pooling quantization process are reduced. The proposed brake pad detection method achieves an average precision of 98.42% and an FPS (frames per second) of 27.77%.
- 上一篇: 高铁物流运输模式及其可行性
- 下一篇: 低空经济与城市轨道交通融合新趋势
过刊浏览
资料下载
友情链接