機械手臂錯誤偵測與診斷

Abstract

在機械手臂嚴重損壞前，事先發覺嚴重錯誤並診斷其錯誤並適時修正改善之，對機械手臂運作壽命而言非常重要。並藉由ISO 9283:1998法規所規定內容分析機械手臂。 為了精確地檢測製程變化和機器人健康指標，本論文透過四軸與六軸機械手臂感測器數值包含各角度、速度與力矩。本論文依ISO 9283:1998所定義指標作為主要實驗指定任務。主要取自其二項評估指標來計算:位姿準確度(Accuracy pose)。其中位姿準確度性可作為長短期記憶模型之重要特徵值，能評估此機械手臂性能藉此訂定與估測健康指數。 藉著ISO 9283:1998 指定的運作方形軌跡下，並以統計製程管制圖(Hotelling T^2 chart)方法監測機械手臂上感測器數據。而後，透過主成分析把機械手臂訊號降低維度並將位姿準確度(Accuracy pose)當作長短期記憶之自編碼模型的重要特徵來訓練模型。隨後，根據多元高斯分布計算機器手臂的異常分數。在根據操作者特性曲線下面積方法等評價模型好壞。最後，並根據異常分數值得出機械手臂健康指標曲線。最後，呈現診斷實驗，探究實驗驗證模型結果。

Obtaining the diagnostic strategy before the catastrophic breakdown is prominent to the industrial robotic arm. It is critical for the robot arm to fix the problem by launching the maintenance strategy at right the time. Testing the performance characteristics of manipulating industrial robots in accordance with ISO 9283. In this thesis, the main results of the experimental work performed on the industrial robotic arm for effective analysis, which includes the position, the torque, and speed from the robotic encoder sensor feedback. Based on the ISO 9283:1998, the calculation of accuracy pose (AP) is the main performance metrics for robot arm. Analyzing the motion of the robot arm under the specific test geometry and the monitor by Hotelling T^2 conrol chart. Then, we modify motion signal by PCA and the AP as prominent feature for the long short term memory neural network based autoencoder scheme. Based on the Gaussian distribution, deliver the anomaly scores for the motion signal subsequently. The model performance metrics is reported by area under the receiver operating characteristics (ROC) curve. Lastly, the robotic health indicator curve is drawn based on the anomaly scores and experimental results are presented and discussed.