作为高铁牵引供电系统的重要组成部分,接触网系统承担着向动车组传输电能的重要功能.实际工程运营表明,受弓网交互产生的持续冲击以及外部环境的影响,接触网支撑部件可能会出现“松、脱、断、裂”等缺陷,导致接触网结构可靠性下降,严重...作为高铁牵引供电系统的重要组成部分,接触网系统承担着向动车组传输电能的重要功能.实际工程运营表明,受弓网交互产生的持续冲击以及外部环境的影响,接触网支撑部件可能会出现“松、脱、断、裂”等缺陷,导致接触网结构可靠性下降,严重影响接触网系统稳定运行.因此,及时精确定位接触网支撑部件(CSCs),对保障高铁安全运行和完善接触网检修维护策略具有重大意义.然而,CSCs的检测通常面临着零部件种类多、尺度差异大、部分零部件微小的问题.针对以上问题,本文提出一种基于多尺度融合金字塔焦点网络的接触网零部件检测算法,将平衡模块和特征金字塔模块相结合,提高对小目标的检测性能.首先,设计了可分离残差金字塔聚合模块(SRPAM),用于优化模型多尺度特征提取能力、扩大感受野,缓解CSCs检测的多尺度问题;其次,设计了一种基于平衡特征金字塔的路径聚合网络(PA-BFPN),用于提升跨层特征融合效率和小目标检测性能.最后,通过对比试验、可视化实验和消融实验证明了所提方法的有效性和优越性.其中,所提的MFPFCOS在CSCs数据集上的检测精度(mAP)能够在达到48.6%的同时,实现30的FLOPs(Floating point operations per second),表明所提方法能够在检测精度和检测速度之间保持良好的平衡.展开更多
The characteristics of chicken droppings are closely related to the health of chickens.Veterinarians often judge the health of a chicken by looking at whether the chicken poop is normal.At present,the inspection of ab...The characteristics of chicken droppings are closely related to the health of chickens.Veterinarians often judge the health of a chicken by looking at whether the chicken poop is normal.At present,the inspection of abnormal chicken droppings in chicken coops relies on manual observation,which is inefficient,accurate varies from person to person,labor-intensive,and has the risk of cross-infection.To achieve efficient,accurate,and intelligent identification of abnormal chicken droppings,an abnormal chicken droppings detection method based on improved Faster Region-based Convolutional Neural Network(Faster R-CNN)was proposed in this study.In the feature extraction network stage,deformable convolution was used and combined with Path Augmentation-Feature Pyramid Network(PA-FPN)to improve the extraction ability of features at different scales.In the Region Proposal Network(RPN)stage,the K-means++algorithm was used to cluster the dataset and obtain the Anchor-ratio which is more suitable for the chicken poop object,and the FocalLoss classification loss function was used to improve the classification ability of difficult samples.In the regional convolutional network stage,the region of interest calibration algorithm was used instead to obtain more accurate localization information.The experimental results show that the improved Faster R-CNN structure can reach an accuracy of 98.8%for abnormal chicken poop detection,and the average accuracy mean value was improved by 27.8%.The results can provide a key core technology support for establishing an efficient abnormal chicken droppings online detection system.展开更多
文摘作为高铁牵引供电系统的重要组成部分,接触网系统承担着向动车组传输电能的重要功能.实际工程运营表明,受弓网交互产生的持续冲击以及外部环境的影响,接触网支撑部件可能会出现“松、脱、断、裂”等缺陷,导致接触网结构可靠性下降,严重影响接触网系统稳定运行.因此,及时精确定位接触网支撑部件(CSCs),对保障高铁安全运行和完善接触网检修维护策略具有重大意义.然而,CSCs的检测通常面临着零部件种类多、尺度差异大、部分零部件微小的问题.针对以上问题,本文提出一种基于多尺度融合金字塔焦点网络的接触网零部件检测算法,将平衡模块和特征金字塔模块相结合,提高对小目标的检测性能.首先,设计了可分离残差金字塔聚合模块(SRPAM),用于优化模型多尺度特征提取能力、扩大感受野,缓解CSCs检测的多尺度问题;其次,设计了一种基于平衡特征金字塔的路径聚合网络(PA-BFPN),用于提升跨层特征融合效率和小目标检测性能.最后,通过对比试验、可视化实验和消融实验证明了所提方法的有效性和优越性.其中,所提的MFPFCOS在CSCs数据集上的检测精度(mAP)能够在达到48.6%的同时,实现30的FLOPs(Floating point operations per second),表明所提方法能够在检测精度和检测速度之间保持良好的平衡.
基金supported by the 'Zhu Cheng Professorial Workstation of China Agricultural University',and the Science and Technology Cooperation Project between China and the United Kingdom under the National Key Research and Development Program(Envirobot an Autonomous Roving Platform for Environment,Health,and Welfare Monitoring of Poultry(Grant No.2017YFE0122200).
文摘The characteristics of chicken droppings are closely related to the health of chickens.Veterinarians often judge the health of a chicken by looking at whether the chicken poop is normal.At present,the inspection of abnormal chicken droppings in chicken coops relies on manual observation,which is inefficient,accurate varies from person to person,labor-intensive,and has the risk of cross-infection.To achieve efficient,accurate,and intelligent identification of abnormal chicken droppings,an abnormal chicken droppings detection method based on improved Faster Region-based Convolutional Neural Network(Faster R-CNN)was proposed in this study.In the feature extraction network stage,deformable convolution was used and combined with Path Augmentation-Feature Pyramid Network(PA-FPN)to improve the extraction ability of features at different scales.In the Region Proposal Network(RPN)stage,the K-means++algorithm was used to cluster the dataset and obtain the Anchor-ratio which is more suitable for the chicken poop object,and the FocalLoss classification loss function was used to improve the classification ability of difficult samples.In the regional convolutional network stage,the region of interest calibration algorithm was used instead to obtain more accurate localization information.The experimental results show that the improved Faster R-CNN structure can reach an accuracy of 98.8%for abnormal chicken poop detection,and the average accuracy mean value was improved by 27.8%.The results can provide a key core technology support for establishing an efficient abnormal chicken droppings online detection system.
