本計畫探討工業4.0智慧工廠(smart factory)之多類別不平衡深度學習(multi-class imbalanced deep learning)嚴重異常徵兆預判(severe anomaly prognosis)技術。我們使用卷積神經網路(convolutional neural network, CNN)搭配長短期記憶(long short-term memory, LSTM)遞歸神經網路(recurrent neural networks, RNN)，以判別出智慧工廠機台是否可能即將發生嚴重異常。所謂嚴重異常(severe anomaly)並不經常發生，但是一旦發生就會對系統產生重大影響，因此需要發展精確的方法加以預測。我們倚賴一些多類別不平衡問題的解決方案以及深度學習強大的分類能力，期待所開發的嚴重異常徵兆預判技術具有良好的效能(也就是高準確度與低計算成本)。我們並且規劃以線切割放電加工機(wire electrical discharging machine, WEDM)智慧工廠作為應用場域，驗證所開發技術的效能。例如，WEDM常見的嚴重異常狀況有線切割銅線斷線、撞機(電擊撞擊加工件)等。銅線斷線後生產會中止，需要重新穿線後繼續生產，會影響生產進度及產品品質穩定度。撞機後機台零件可能會有所損毀，造成生產立即中止，而且會產生零件部位移位等狀況，需要等待長時間之重新調校才能繼續生產，影響甚鉅。我們必須精準的預測嚴重異常的發生，以積極進行適當的主動(proactive)維護規劃。 ;The project investigates multi-class imbalanced deep learning for severe anomaly prognosis in Industry 4.0 smart factories. We apply convolutional neural network (CNN) along with recurrent neural networks (RNN) of the long short-term memory (LSTM) model to determine if severe anomalies may appear in a smart factory machine. The so-called severe anomaly happens rarely, but once it happens, it usually has a significant impact on the system. Therefore, accurate methods are needed to predict its happening. We rely on a number of solutions to multi-class imbalanced deep learning and expect that the developed severe anomaly prognosis technology has good performance. We also plan to apply the developed technology to the wire electrical discharging machine (WEDM) smart factories to validate the performance of the developed technology.
