基于多尺度融合金字塔焦点网络的接触网零部件检测

A detector based on a multiscale fusion pyramid focus network for catenary support components

  • 摘要: 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MFP-FCOS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.

     

    Abstract: As a crucial component of a high-speed rail traction power supply system, the catenary system is responsible for transmitting electrical energy to electric multiple units (EMUs). In practice, continuous impacts from pantograph-net interactions and external environmental factors can lead to defects in the catenary’s supporting parts, such as looseness, detachment, fracture, and cracking. These issues compromise the reliability of the catenary structure and pose risks to its stable operation. Therefore, timely and accurate positioning of the catenary support components (CSCs) is vital for ensuring the safe operation of high-speed rails and improving the catenary maintenance strategies. In 2012, the former Ministry of Railways of China (now the China Railway Corporation) officially promulgated the General Technical Specifications for High-speed Railway Power Supply Safety Detection and Monitoring System. This study marked a shift from traditional manual inspection methods to intelligent non-contact catenary detection and maintenance using computer vision technology. This study addresses challenges in detection systems by focusing on “catenary part positioning” in the whole detection process from the perspective of the functional integrity of the detection system. Detecting CSCs is challenging because of the variety of parts, scale differences, and small size of components. To overcome these challenges, this study proposes a catenary component detection algorithm that utilizes a multiscale fusion pyramid focus network. This approach integrates a balance module and a feature pyramid module to improve the detection performance of small targets. The separable residual pyramid aggregation module (SRPAM) was designed to optimize multi-scale feature extraction, expand the receptive field, and address multi-scale issues in CSC detection. Furthermore, a path aggregation network based on the equilibrium feature pyramid (PA-BFPN) was designed to improve cross-layer feature fusion efficiency and small object detection performance. Finally, the effectiveness of the proposed method is demonstrated through comparative experiments, visual analysis of the results, multi-scale feature fusion module experiments, feature pyramid network experiments, and ablation studies. The results demonstrate that the proposed multiscale feature pyramid FCOS (MFP-FCOS) algorithm offers excellent overall performance compared to many classical algorithms. Visualization experiments confirm its effectiveness in detecting targets across different scales and effectively solving small-scale and multi-scale sample detection challenges. The proposed SPRAM effectively mitigates information loss and improves feature extraction performance, whereas the proposed PA-BFPN obtains more comprehensive feature information. In summary, the proposed MFP-FCOS achieved a detection accuracy (mAP) of 48.6% on the CSC dataset with 30 floating point operations per second (FLOPs), indicating a balanced trade-off between detection accuracy and detection speed.

     

/

返回文章
返回