Authors

Abstract

To achieve precise prevention and control of rubber roll wear in huskers, this study established a complete experimental platform integrating mechanical transmission, feeding control, and visual acquisition units, and subsequently developed a real-time monitoring system for paddy flow instability based on machine vision. An image acquisition system composed of a CMOS camera and a customized light source was built to construct a dedicated dataset for paddy flow states. Based on the lightweight YOLOv8n detection model, Python programming was adopted in the PyCharm environment with the ultralytics library integrated, realizing real-time recognition and quantitative analysis of paddy flow states. The results demonstrated that the system realized a real-time detection efficiency of 14 FPS on a local workstation, and the YOLOv8n model achieved a recognition accuracy of 90.8% in terms of mean Average Precision at IoU threshold 0.5 (mAP@0.5) for sparse and overlapping grain states. The system could effectively capture key abnormal states, including inclined paddy grains entering the rolling zone, sparse paddy flow lasting more than 5 seconds, and overlapping paddy flow density exceeding 10 grains/cm². This study transformed the mechanical characteristics of paddy flow instability into pixel-level quantitative indicators and established an integrated visual monitoring paradigm of "perception-analysis-decision", providing effective technical support for the intelligent management and control of huskers.

Abstract in Chinese

为实现砻谷机胶辊磨损的精准防控，本研究搭建了集成机械传动、喂料控制与视觉采集单元的完整实验平台，开发了基于机器视觉的稻谷料流失稳实时监测系统。采用CMOS相机与定制光源构建图像采集系统，构建稻谷料流状态专用数据集；基于YOLOv8n轻量化检测模型，在PyCharm环境中采用Python编程集成 ultralytics库，实现料流状态的实时识别与量化解析。结果表明，该系统在本地工作站可实现14FPS的实时检测效率，YOLOv8n模型对颗粒稀疏与堆叠状态的识别精度（交并比0.5阈值下的平均精度均值，mAP@0.5）达到90.8%，可有效捕捉稻谷颗粒入轧倾斜、料流稀疏持续时间大于5秒、料流堆叠密度超过10粒/cm²等关键异常状态。本研究将料流失稳的力学特征转化为像素级量化指标，构建“感知-解析-决策”一体化视觉监测范式，为砻谷机智能管控提供了有效的技术支撑。