| 摘要: | 近年來由於資訊科技的快速發展，帶動全球許多產業朝向數位化轉型，許多傳統製造業也逐步發展工業4.0(Industry 4.0)智慧工廠。這些工廠希望使用物聯網(Internet of Things, IoT)、大數據(big data)、機器學習(machine learning)以及雲端計算(cloud computing)等先進技術，實現智慧製造(smart manufacturing)的目標。如此一來，製造業對於生產系統上的產品生產品質能有更好的管控，達成更高的良率。本論文聚焦於線切割放電加工(wire electrical discharge machining, WEDM)產品的表面粗糙度(surface roughness)預測，使用遷移學習(transfer learning)技術於兩種不同工件材質，其中一個是資料數量較多的材質A工件，在此視為來源域(source domain)；而另一個則是資料數量較少的材質B工件，在此視為目標域(target domain)。
具體而言，本論文先藉由數量較多的材質A工件資料，在加工之前透過靜態生產參數，並在加工之後透過靜態生產參數以及動態機台狀態資料，訓練出可以預測工件表面粗糙度的神經網路模型，包括深度神經網路(deep neural network, DNN)以及門控循環單元(gate recurrent unit, GRU)與深度神經網路結合的神經網路預測模型。然後使用兩種遷移學習作法：第一種為權重鎖定(weight-freezing)方法，藉由數量較少的材質B工件資料，將原先提供訓練好可以預測材質A工件表面粗糙度的神經網路模型，轉換為可以預測材質B工件表面粗糙度的神經網路模型。第二種遷移學習為多任務學習(multitask learning)：藉由共同使用數量較多的材質A工件資料以及數量較少的材質B工件資料，同時訓練出可以預測材質A以及材質B工件表面粗糙度的個別神經網路模型。實驗結果顯示，二種遷移學習方法都可以充分利用數量較少的目標域資料，有效提升來源域與目標域的神經網路模型的預測準確度。;With the rapid development of information technology in recent years, many industries around the world are moving toward digital transformation, and many traditional manufacturing industries are also gradually developing Industry 4.0 smart factories. The factories intend to utilize advanced techniques such as the Internet of Things (IoT), big data, machine learning, and cloud computing, to reach the goal of smart manufacturing. In this manner, the product quality is better controlled and the yield rate is improved. This paper focuses on the prediction of the surface roughness (SR) of wire electrical discharge machining (WEDM) workpieces with transfer learning techniques for two different materials, denoted as material-A and material-B. The data of material-A workpiece are regarded as source domain data, whose amount is larger, whereas the data of material-B workpiece are regarded as target domain data, whose amount is smaller.
First, material-A workpiece data are applied to train neural network models for SR prediction. Specifically, static manufacturing parameters are used to train deep neural networks (DNNs) for SR prediction before manufacturing, whereas static manufacturing parameters along with dynamic manufacturing conditions are used to train gate recurrent unit (GRU) networks for SR prediction after manufacturing. Afterwards, two transfer learning methods are utilized. The first method is weight-freezing. It uses a small amount of material-B workpiece data to train a neural network model that can predict surface roughness of material-B workpiece on the basis of well-trained models that can predict surface roughness of material-A workpiece. The second method is multi-task learning. It uses both material-A workpiece data, whose amount is larger, and material-B workpiece data, whose amount is smaller, to train separate neural network models for SR prediction for material-A and material-B workpieces, respectively. The experimental results show that the two transfer learning methods both can efficiently improve prediction accuracy of source domain and target domain neural network models through using only a small amount of target domain data. |
