多數量產化的裝置和產品皆是以自動化的組裝機台或機器手臂進行組裝和製造，但仍有少部分的產品組裝和製造過程無法使用機器執行量產，而手動組裝常常遇到使用工具施力不當或施力角度不對導致產品良率不佳。本論文提出透過收集動作數據，設計一套智慧型螺絲起子辨識系統，該系統透過動作辨識的方式針對產線上人員執行產品組裝產生失誤執行糾錯，人員在執行產品裝配時會按造固定SOP執行操作，本文將螺絲起子常用到的動作分為六大類(1)往下鎖螺絲(2)將手提起(3)往下平推螺絲(4)往左平推螺絲(5)往上平推螺絲(6)往右平推螺絲，產線上組裝的SOP會由此六大類動作組合而成。將採集動作訊息則是透過MetaMotionC慣性感測器執行動作採樣，並將採樣的數據輸入智慧型螺絲起子辨識系統，輸入的資料會先使用感測器歸零校正模組進行校正，並將校正後的參數輸入至四元數轉換模組執行姿態預測並融合為四元數輸出給運動特徵模組，由運動特徵模組執行特徵點辨識和動作分類，測試者需以標準動作以供實驗樣本建立資料庫，實驗成果根據神經網路機器學習的評估指標採用的成效衡量指標來衡量優劣。 實驗發現使用智慧型螺絲起子辨識系統的動作辨識模組內部1D CNN Model測試最高可以得到78.9%的辨識率，動作辨識模組內部LSTM Model經測試最高可以得到66.7%的辨識率，若未使用智慧型螺絲起子辨識系統直接執行1D CNN神經網路得到的辨識為26.3%、LSTM神經網路得到的辨識為26.5%，可以發現直接輸入動作數據給神經網路會因為數據雜訊導致神經網路判斷錯誤。 使用智慧型螺絲起子辨識系統的相較於沒有使用智慧型螺絲起子辨識系統的辨識率1D CNN Model提升了52.6%辨識率、LSTM Model提升了52.6%辨識率，智慧型螺絲起子辨識系統的1D CNN Model辨識率最高78.9%，實驗證明透過本系統實行對作辨識將有效的達到減少作業人員裝配疏失的需求，進而提升產品良率。 ;Many mass-produced devices and products are assembled and manufactured by automated assembly machines or robotic arms, but there are still a small number of product assembly and manufacturing processes that cannot be mass-produced by machines, and manual assembly often encounters the use of tools Improper application of force or the wrong angle of application leads to poor product yield.This paper proposes to design a smart screwdriver identification system by collecting motion data. The system uses motion recognition to correct errors in product assembly performed by personnel on the production line. When personnel perform product assembly, they will perform fixed SOP implementation. Operation, this article divides the commonly used actions of the screwdriver into six categories (1) lock the screw down (2) lift the hand (3) push the screw down (4) push the screw to the left (5) push the screw up and flat (6) Push the screw horizontally to the right, and the SOP assembled on the production line will be composed of six types of actions.The collected motion information is to perform motion sampling through the MetaMotionC inertial sensor, and input the sampled data into the intelligent screwdriver identification system. The input data will be calibrated by the sensor zero calibration module first, and then the calibration will be performed. The parameters are input to the quaternion conversion module to perform posture prediction and merge into a quaternion and output to the motion feature module. The motion feature module performs feature point identification and action classification. The tester needs to take standard actions for the establishment of experimental samples In the database, the experimental results are used to measure the pros and cons of the results according to the evaluation indicators of neural network machine learning. The experiment found that the 1D CNN Model test inside the action recognition module using the smart screwdriver recognition system can get the highest recognition rate of 78.9%, and the LSTM Model inside the action recognition module can get the highest recognition rate of 66.7% after the test. If wisdom is not used The type screwdriver identification system directly executes the 1D CNN neural network to obtain an identification of 26.3%, and the LSTM neural network to obtain an identification of 26.5%. It can be found that directly inputting the action data to the neural network will cause the neural network to judge due to data noise. mistake. Compared with the recognition rate of the smart screwdriver recognition system that uses the smart screwdriver recognition system, the recognition rate of the 1D CNN Model increased by 52.6%, the LSTM Model increased the recognition rate by 52.6%, and the 1D CNN of the smart screwdriver recognition system The model recognition rate is as high as 78.9%. Experiments have proved that the implementation of this system will effectively reduce the need for assembly errors by operators, thereby increasing the product yield.
