Evaluating the process operating state taking into consideration operator interventions with application to a hot rolling mill process

In complex industrial processes, process operators often intervene in automatic control systems based on their assessment of the operating state. Traditional operating state evaluation methods do not take into consideration or cannot effectively use this intervention information, and thus may incorrectly evaluate the operating state. In this paper, a convolutional-neural network-conditional variational auto-encoder (CNN-CVAE)-based method for evaluating the operating state is proposed to address this problem. First, the operating-state labels are constructed considering the operator-intervention information. Next, the features of operator-intervention variables (OIVs) are extracted based on CNN, and the obtained probabilities of belonging to different operating states are used as conditional probabilities of CVAE to supervise the feature extraction from the ordinary process data. Finally, both features are fused in a fully connected layer to obtain the predicted operating state. Compared with traditional methods, CNN-CVAE can capture features from both OIVs and process data for evaluating the operating state. The proposed method is validated in a real, hot strip rolling mill process. The results show that the proposed method improves the evaluation accuracy by 54.72% compared with five methods that do not fully use the OIVs.