利用循環神經網路之製程預測控制最佳化

Abstract

本研究以數位分身(Digital Twin)概念出發，以深度神經網路模型作為虛擬軋機模擬實際軋延系統，由於在軋延製程受到機器參數、控制參數、鋼板特性、過往軋延結果等因子影響，無法直接以物理或數學模型進行分析、實驗，我們將軋機內部視為黑盒子，建立AI模型作為虛擬系統，經由過往蒐集的大量數據，學習機器特性、了解出入口關係，進而精準預測出口板形。 現行軋機由內漸模糊控制系統控制，時常發生控制不到位的問題，我們結合虛擬軋機與模擬最佳化模型計算出在不同軋延情況及目標下最適合的控制參數，使得最終出口板形接近目標板形，提升製程品質及效率。 考量控制參數由油壓輥控制，作動速度緩慢，加上模擬最佳化模型計算耗時，不符合實時控制的需求。在蒐集足夠多資料後，我們建立基於長短期神經網路的編碼器-解碼器模型，學習面對不同入口板形及目標板形的最佳控制方式，省去模擬最佳化計算時間，並可以在偵測到入口板形後立即預測未來多步的最佳控制方式，使控制器能提早移動最佳控制位置，利用模型預測控制方法提升實際產線的實作可行性並優化製程品質。

Based on the concept of Digital Twin, we build a deep neural network as a virtual rolling mill to simulate the operation of reversing cold rolling mill. The rolling process is affected by machine parameters, control parameters, raw materials, previous rolling, and so on. Therefore, the rolling process can’t be interpreted by mathematical or physical models. We assume that the interaction of the rolling mill is a black box and establish an AI model as a virtual system which can precisely predict the shape after rolling. The current control method of the reversing cold rolling mill is based on the built-in fuzzy control system, which often results in a gap between the exit shape and the target shape. Based on the virtual rolling mill, simulation optimization model to calculate the appropriate control parameters under different target shape and environments, which optimize the output shape and improve the process quality. Consider that the control parameters are controlled by hydraulic controllers which moves slowly compared with the strip and that simulation optimization model takes long time on calculation. We establish LSTM based Encoder-Decoder model to predict multi-steps ahead optimal control actions. Consequently, the controllers can move to optimized position. The results for the neural network model predictive control (NNMPC) overall show better performance in the control of the system over the original one.