- 學位論文
以機器學習動力學模型建構外扭力觀測器和虛擬力規
Developing External Torque Observers and a Virtual Force Sensor Using Machine Learning Dynamic Models
摘要
不論為插件任務或洞棒組裝策略的研究上，大多文獻皆會透過外部安裝六軸力規或壓力感測器以達到力的控制與判斷，然而近幾年機器手臂通常具備有教導模式，代表手臂自身各軸具有可以量測扭力值的感測器，如果可以善加利用其設計虛擬力規則可以省去外加感測器之成本與維護。 本研究比較了常見的DNN、LSTM、XGBoost和LightGBM機器學習方法於學習六軸機械手臂之動力學模型的性能，最後設計與分析虛擬力規。為了達成此目標其過程包含了訓練軌跡生成演算法、動畫模擬、實機收集訓練資料、透過超參數最佳化方法提升機器學習性能、設計外扭矩觀測器、推導及生成虛擬力規、推導力控制器(流程控制、PID、阻抗控制)，最後透過三個實際應用驗證其可行性。 為了快速生成大量與全域的訓練軌跡，設計了訓練軌跡生成演算法，其包含了組態生成、工作範圍檢查、自身桿件碰撞檢查與排列組態四個步驟，最後再經由MATLAB Robotics Toolbox預先模擬其實際移動過程。 在動力學模型比較的部分，比較四種型態動力學模型、不同訓練資料量其效果，並比較了不同的機器學習方法、搭配超參數最佳化方法調整之性能。測試資料分析上，手臂前三軸模型扣除真實扭力其與數學模型扣除真實扭力之均方根誤差(root-mean-square error, RMSE)值，而第四軸、第五軸則透過真實力規實際分析其在實際力控制上的力追蹤之準確性並透過定點力追蹤分析其線性回歸之斜率以確保模型正確。 在外扭矩觀測器上透過文獻常見的動力學方程式求得外扭矩，並推導Jacobian得到機械手臂關節扭矩與端效器受力之間的關係，再藉由外扭矩觀測器作為特徵、WACOH力規值作為標籤透過機器學習方法訓練出追隨手臂端效器姿態的Approach vector方向上的虛擬力規。 最後透過三個實驗證明其可行性，並透過WACOH六軸力規比較與分析其實際應用上的精確度。 虛擬力規的精確性自然是不如外部安裝六軸力規來得準確，但透過本研究的實驗與所提出之洞棒組裝策略說明其可行於擦桌子、擦玻璃、具有一定程度誤差之洞棒組裝等，不需要非常精確力的控制任務上。
並列摘要
Regardless of the research on inserting components or peg in hole assembly strategies, most of the literature will use externally installed six-axis force sensor or pressure sensors to achieve force control and judgment. However, in recent years, the robot arm usually has a teaching mode, which means that each axis of the arm has a sensor that can measure the torque value. If this can be effectively used to design virtual force sensor, the cost and maintenance of additional sensors can be saved. This study compares the performance of DNN, LSTM, XGBoost, and LightGBM usual machine learning methods to learn the dynamic model of a six-axis robotic arm, and finally designs and analyzes virtual force sensor. In order to achieve this goal, the process includes training trajectory generation algorithm, animation simulation, real machine collection of training data, improvement of machine learning performance through hyperparameter optimization method, design of external torque observer, derivation and generation of virtual force sensor, derivation force controller (process control, PID, impedance controller), verified through three applications. In order to quickly generate a large number of global training trajectories, a training trajectory generation algorithm is designed, which includes four steps: configuration generation, workspace check, self-collision check and Permutation configuration. Finally, it is pre-simulated by MATLAB Robotics Toolbox. In the comparison of dynamics models, the effects of four types of dynamics models and different amounts of training data are compared, and the performance of different machine learning methods and hyperparameter optimization methods are compared. In the analysis of test data, comparing the respective RMSE(root-mean-square error) values of the three axis model of the arm minus the real torque and the mathematical model minus the real torque, and the fourth axis and the fifth axis are actually analyzed the accuracy of the force tracking on the actual force control by the real force sensor. And through the fixed-point force tracking analyzes the slope of its linear regression to ensure that the model is correct. On the external torque observer, the external torque is obtained through the usual dynamic equations in the literature, and Jacobian is derived to obtain the relationship between the joint torque of the robot arm and the force of the end effector. Then through machine learning methods use the external torque observer as a feature and force value of WACOH force sensor as a label to train a virtual force sensor model in the direction of Approach Vector that follows the posture of the arm end effector. Finally, through three experiments to prove its feasibility, and through WACOH six-axis force sensor to compare and analyze the accuracy of its practical application. The accuracy of the virtual force sensor is intuitively less accurate than the external installation of a six-axis force sensor. However, through the experiments of this study and the proposed peg in hole assembly strategy, it is feasible to wipe the table, wipe the glass, and peg in hole assembly with some degree error. It is feasible for any control task that does not require very precise force.