针对常规深度学习网络规模大、对现场设备硬件要求高且人工标注位置数据复杂费时等问题,提出了一种语义数据标注的轻量化轨道扣件故障检测方法.该方法仅对训练数据做语义标注,改进轻量化Transformer模型,嵌入梯度加权类激活映射(gradient-weighted class activation mapping,简称Grad-CAM)模块对模型输出的特征图权重作映射处理,可将模型对轨道扣件检测效果可视化.将获得的激活图进行二值化定位检测目标位置,实验结果表明,在真实铁路环境下,改进的轻量化轨道扣件模型的准确率为94.31％.
Abstract
In response to the challenges of the large scale of conventional deep learning networks,high hard-ware demands for field devices,and the intricate and labor-intensive process of manually annotating location data,this paper introduces an innovative approach to rail fastener fault detection.This method only makes se-mantic annotation for training data,improves the lightweight Transformer model.Further,incorporating the gradient-weighted class activation mapping(Grad-CAM)module can visualize the weight distribution of the fea-ture map output generated by the model.This visualization provides insights into the model performance in rail fastener detection.Subsequently,the resulting activation map is binarized to precisely pinpoint and identify the target location.Experimental results demonstrate that the improved lightweight model for track fastener detec-tion achieves an impressive accuracy rate of 94.31％in real-world railway environments.
维普
万方数据