基于关键点识别的转辙机动静接点组安装效果检测方法

Detection Method for Installation Effect of Switch Machine Movable/Static Contact Groups Based on Key Point Recognition

[目的]对转辙机动静接点组安装效果的传统检测方式存在速度慢、精度差、容易受到人为因素影响等问题。有必要引入快速、高精度的图像检测技术,基于关键点识别,对转辙机动静接点组安装效果检测方法进行研究。[方法]详细描述了转辙机动静接点组安装效果检测方法的检测过程。介绍了转辙机动静接点组的24个识别关键点,阐述了动接点接触深度和底座间距的计算方法。通过不同组合模型的测试分析,所选择的关键点识别最优模型以YOLOv8视觉框架中的姿态检测算法作为基础,引入了BiFormer双编码器注意力机制和SCConv高效卷积模块,并描述了拍摄辅助框和透视矫正变换功能。[结果及结论]关键点识别最优模型的识别用时仅1.3 ms,识别正确率达到了96.3%。基于关键点识别的转辙机动静接点组安装效果检测方法对动静接点组的图像识别率达到99.8%,计算精度达±0.1 mm,识别平均误差不超过0.3%,每组的检测用时仅2 s。可见,该方法远比人工检测方法更智能、高效。

[Objective]Traditional methods for detecting the installation effect of SmMSC(switch machine movable/static contact)groups are slow,inaccurate,and susceptible to human errors.It is necessary to introduce a fast and high-precision image detection technology based on key point recognition,aiming to make research on the installation effect detection method for SmMSC groups.[Method]The detection process of the above-mentioned detection method is described in detail.24 key recognition points of the SmMSC groups are introduced,along with the calculation method for the contact depth of moving contacts and spacing between the bases.Through test analysis of different combination models,the selected optimal key point recognition model is based on the pose detection algorithm in the YOLOv8 visual framework,incorporating the BiFormer dual-encoder attention mechanism and SCConv(spatial and channel reconstruction convolution)efficient convolution module.The functions of the auxiliary shooting frame and perspective correction transformation are also described.[Result&Conclusion]Recognition time of the optimal key point recognition model is only 1.3 milliseconds,with a recognition accuracy reached 96.3%.The installation effect inspection method of SmMSC groups based on key point recognition achieves an image recognition rate of 99.8%for the dynamic and static contact groups,with a calculation accuracy of±0.1 mm,and an average recognition error lower than 0.3%,only 2 seconds for each group′s detection.It′s obvious that this method is significantly more intelligent and efficient compared to manual detection methods.