| 摘要: | 在工廠量產產品時，依據產品規格與產量指派產線作業員進行產品裝配，但由於產
線員工裝配產品時，易發生人為操作失誤導致產品破裂或是產品漏裝，進而導致影響整
體產品出貨品質與製造成本。
故本篇論文提出以作業員戴上指套且指套上黏附九軸感測器和手握自動螺絲起子，
蒐集動作感測數據，再以卡爾曼濾波進行感測數據濾波，以越零率方式進行動作切割，
再利用Mahony 濾波計算誤差模型來校正感測器數據後計算出四元數，利用四元數當作
特徵後利用一維卷積神經網路來分類 (1) 右上頂螺絲、(2) 右下頂螺絲、(3) 右斜上頂
螺絲、(4) 左斜下頂螺絲、(5) 向下鎖螺絲、(6) 抬手放下，共六種作業員裝配基本動作，
由此六種裝配動作組合順序以判別裝配產品是否正確。
實驗結果發現利用Mahony 濾波校正感測器後再計算出四元數為特徵進行一維卷積
神經網路1DCNN 能提升72% 的辨識率，進行長短期記憶模型LSTM 能提升30%辨識
率，進行CNN-LSTM 也能提升70%的辨識率，而校正後的數據在1DCNN 有高達94%
的平均辨識率，比使用LSTM 模型高39%辨識率，比使用CNN-LSTM 高3%辨識率，
故使用此方法能預防現今作業員鎖螺絲動作出錯與降低作業員動作訓練成本，與改善產
線員工管理智能化不足等問題。;When mass-producing products in the factory, the production line operators are assigned
to carry out product assembly based on product specifications and product number. However,
when the production line employees assemble products, human operation errors are prone to
cause product rupture or product miss screw , which in turn affects the overall product quality
and manufacturing cost.
Therefore, This paper proposes that the operator wears a finger sleeve with a nine-axis
sensor attached and holds an automatic screwdriver to collect motion sensing data, and then
uses the Kalman filter to filter the sensing data, and performs motion cutting with a zerocrossing
rate. Then use Mahony filter to calculate the error model to correct the sensor data and
calculate the quaternion, use the quaternion as a feature, and then use a one-dimensional
Convolutional neural network to classify (1) upper right screw , (2) lower right screw ,(3) right
diagonal screw,(4) Left diagonal screw,(5) down lock screw ,(6) Raise hand and lay down , total
of six basic assembly actions for operators, The sequence of six assembly actions is used to
judge whether the assembled product is correct.
The experimental results found that the use of Mahony filter to calibrate the sensor and
then calculate the quaternion as the feature for one-dimensional convolutional neural
network 1DCNN can increase the recognition rate by 72%, and the long- and short-term
memory model LSTM can increase the recognition rate by 30%. CNN-LSTM can also increase
the recognition rate by 70%, and the corrected data has an average recognition rate of 94%
in 1DCNN, which is 39% higher than the LSTM model and 3% higher than the CNN-LSTM.
Using this method can prevent the current operator from locking screws and reduce the cost
of operator training, and improve the lack of intelligent management of production line
employees. |
