應用光定位技術於機械手臂校正之系統開發;Development of a Robotic Arm Calibration System Using Light Positioning Technology

NCU Institutional Repository > 工學院 > 機械工程研究所 > 博碩士論文 > Item 987654321/98657

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98657

题名:	應用光定位技術於機械手臂校正之系統開發;Development of a Robotic Arm Calibration System Using Light Positioning Technology
作者:	蔡劭傑;Tsai, Shao-Chieh
贡献者:	機械工程學系
关键词:	光定位;機器學習;機械手臂;重定位;快速校正;三維定位;Visible Light Positioning;Machine Learning;Robotic Arm;Re-localization;Fast Calibration;3D Positioning
日期:	2025-07-14
上传时间:	2025-10-17 13:03:21 (UTC+8)
出版者:	國立中央大學
摘要:	在現今自動化製造環境中，機械手臂常因場域變動或產線調整需進行重定位與校正。傳統以相機或深度相機為基礎的視覺方案，存在硬體成本高與視野範圍與解析度難以兼顧的限制，難以滿足大範圍高精度校正需求。為解決上述問題，本研究提出一套整合光定位技術與機器學習的創新機械手臂校正方法。系統透過在空間中佈設可調變訊號的光源，並由安裝於機械手臂末端的光學感測器接收訊號，進而建立光訊號與三維座標之間的對應關係。結合訊號處理、特徵篩選與機器學習模型，實現高精度的位置估測與自動化校正。經由大量實驗優化硬體配置並對五種機器學習模型的效能進行了比較，這些模型包括：隨機森林迴歸(Random Forest Regression, RFR)、多層感知器(Multilayer Perceptron, MLP)、Transformer 模型，以及兩種基於柯爾莫哥洛夫-阿諾爾德網絡(Kolmogorov-Arnold Networks , KAN)的架構，其一為高效版本(Efficient- KAN)、另一則為最新版本(KAN 2.0)。實驗結果顯示KAN 2.0表現最佳，整體平均誤差為 4.407 mm，將光定位系統在三維空間的定位誤差控制在毫米等級。此外，實驗亦驗證所提訊號處理與特徵篩選技術具備良好的環境適應性。本研究成功驗證該方法能有效應用於機械手臂校正任務，具備良好的可行性與應用潛力，滿足多變生產場域的自動校正需求。;In today’s automated manufacturing environments, robotic arms frequently need re-localization and calibration due to the operation or production line adjustments. Traditional vision-based solutions relying on high-cost visual or stereo cameras are restricted in the trade-off between achieving a wide field of view and high resolution, resulting in a challenge in meeting the demands of large-scale, high-precision calibration. This study proposes an innovative robotic arm calibration method that integrates light positioning technology with machine learning to address these issues. The system deployed the light sources with the modulated signals. It utilized the optical sensors mounted on the robotic arm’s end-effector to receive these signals, thereby establishing a correspondence between the optical signals and three-dimensional coordinates. High-precision position estimation and automated calibration can be achieved via signal processing, feature selection, and machine learning models. Through extensive experiments optimizing hardware configurations and comparing five models, mainly Random Forest Regression(RFR), Multilayer Perceptron(MLP), Transformer, Efficient-Kolmogorov-Arnold Networks,(Efficient-KAN), and Kolmogorov-Arnold Networks 2.0(KAN 2.0), KAN 2.0 was identified the best performance with an overall mean absolute error (MAE) of 4.407 mm, effectively controlling the spatial positioning error of the light positioning system in a millimeter level. Furthermore, the experiments also verified that the proposed signal processing and feature selection techniques possess good environmental adaptability. In conclusion, the proposed method can be effectively applied to robotic arm calibration tasks, demonstrating good feasibility and application potential, and meeting the automated calibration demands of dynamic production environments.
显示于类别:	[機械工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	12	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....