基于C3F-YOLOv5的轻量化列车车底螺栓检测方法研究

Research on underbody bolt detection method of lightweight train based on C3F-YOLOv5

提高目标检测在列车车底复杂背景下螺栓的识别精度及检测速率,对于提高列车行驶安全性具有重要意义。为高效检测螺栓,提出一种基于C3F-YOLOv5的轻量化列车车底螺栓检测方法。使用自行设计的履带式车底检测机器人获取车底螺栓图片,并混合自行搭建的模拟平台螺栓图片与真实的列车车底螺栓图片作为最终数据集。将C3模块中的Bottleneck结构替换为Faster_Block结构,改进为C3F模块;并分别与FasterNet、GhostNet和MobileNetV3轻量化结构进行对比。此外,引入注意力机制CA模块,同时将原有损失函数L_(GIoU)替换为更为契合的L_(MPDIoU);使用SE模块和CBAM模块分别与CA模块进行比较,作为消融实验。最后采用LAMP分数的计算方法对模型权重参数排序,并将不重要的权重参数剪枝,作为新型模型压缩的方法。将最终模型压缩后的C3F-YOLOv5s网络模型分别与YOLOv4、YOLOv7、Mask R-CNN、RetinaNet进行对比试验。研究结果表明,在使用混合数据集的情况下,最终网络模型的平均检测精度达到了92.8%,检测速度达到了256.7 FPS。相较于其他4种经典的深度学习网络模型,改进后的模型检测效果更好,同时模型表现出较强的鲁棒性和泛化性能。该方法可在无法获得更多真实车底螺栓图片时,使训练后网络更加适应真实车底的情况,改进后的算法可以同时提高螺栓的识别精度和检测速度,可以为后续相关研究提供技术参考和理论支撑。

It is of great significance to improve the identification accuracy and detection speed of bolt under the complex train background.In order to more efficiently detect bolts,a method for detecting bolts under lightweight train based on C3F-YOLOv5 was proposed.A self-designed crawler bottom detection robot was used to obtain the bolt pictures of the train bottom.The bolt pictures of the simulated platform were mixed with the real bolt pictures of the train bottom as the final data set.The Bottleneck structure in the C3 module was replaced with a Faster_Block structure,improving it to the C3F module.Furthermore,the C3F module was compared with the lightweight structures of FasterNet,GhostNet,and MobileNetV3.In addition,the attention mechanism CA module was introduced,and the original loss function LGIoU was replaced with the more suitable LMPDIoU.The ablation experiments were carried out,and the SE module and CBAM module were added to compare with the CA module,respectively.Finally,LAMP score was used to sort the weight parameters of the model,and the unimportant weight parameters were pruned as a new model compression method.The C3F-YOLOv5s network model compressed by the final model was compared with YOLOv4,YOLOv7,Mask R-CNN,and RetinaNet.The research results show that when using mixed datasets,the average detection accuracy and detection speed of the final network model reach 92.8%and 256.7 FPS,respectively.Compared to the other four classic deep learning network models,the improved model exhibits superior detection performance,robustness,and generalization performance.This method can make the network more adaptable to the real situation of the train bottom after training when more pictures of the real train bottom cannot be obtained.The improved algorithm can promote the identification accuracy and detection speed of bolts,and can provide technical reference and theoretical support for subsequent related research.