基于双模态与模糊自适应PID控制的甘蔗收获机切梢器智能定位系统研究

Intelligent Positioning System of Sugarcane Harvester Tip Cutter Based on Bimodal and Fuzzy Adaptive PID Control

摘要:

为解决甘蔗收获机驾驶员依赖肉眼和经验判断蔗梢位置，难以实时精准调整切割高度的问题，提出了一种基于双模态及模糊自适应PID控制的甘蔗收获机切梢器智能定位系统，并提出了基于甘蔗簇特征识别的簇调刀策略。系统首先利用YOLO v8实例分割模型检测蔗梢，深度相机实时采集甘蔗蔗梢的深度数据，并将像素坐标转换为相机坐标，实现高度测量，为验证深度相机在田间的实际定位效果，设计了多组田间甘蔗定位试验，结果显示，深度相机在距离甘蔗50~100cm范围内的平均相对误差为0.189%~0.949%。随后，通过设计模糊规则，单片机结合模糊自适应PID算法，上升响应速度约为289.36mm/s，下降响应速度约为273.16mm/s，误差在±1mm内。PID对比试验表明，模糊PID控制相较于传统PID控制，模糊PID响应时间缩短了0.12s，为1.24s，超调量由12.80mm减少至4.63mm，超调次数降至1次，稳态误差稳定在±2mm以内。最后，动态试验表明，系统的平均识别时间为0.08s，具有良好的实时性。

Abstract:

Aiming to address the challenge faced by sugarcane harvester operators who rely solely on visual judgment and experience to determine cane tops’positions, making it difficult to adjust cutting height accurately and in real-time, an intelligent positioning system for sugarcane harvester top cutters was proposed based on dual-modal and fuzzy adaptive PID control. Additionally, a cluster-based knife adjustment strategy was developed based on sugarcane cluster feature recognition. The system firstly utilized the YOLO v8 instance segmentation model to detect cane tops, with a depth camera capturing depth data of sugarcane tops in real time and converting pixel coordinates into camera coordinates for height measurement. To validate the depth camera’s field performance, multiple field experiments were conducted. Results showed that within a range of 50~100cm from the sugarcane, the average relative error of the depth camera ranged from 0.189% to 0.949%. Subsequently, fuzzy rules were designed, and a microcontroller integrated with a fuzzy adaptive PID algorithm was used to control the servo motor’s operating speed. The maximum rising response speed was approximately 289.36mm/s, and the descending response speed was about 273.16mm, with a measurement error within ±1mm. PID comparison experiments showed that fuzzy PID control, compared with traditional PID control, reduced the response time by 0.12s, bringing it down to 1.24s. The overshoot was decreased from 12.80mm to 4.63mm, the number of overshoots dropped to one, and the steady-state error stabilized within ±2mm. Finally, dynamic tests demonstrated that the system’s average recognition time was 0.08s, showcasing excellent real-time performance.