A Multi-Operation Mode KPI Control Method Based on Regression Rule Extraction from Deep Neural Networks

In process industries, the occurrence of diverse operational states and complex spatio-temporal relationships often complicates the control of key performance indicators (KPIs). Additionally, existing KPI control methods generally lack the ability to balance control accuracy and interpretability, posing significant challenges to the safe and stable control of KPI. In response to these challenges, this paper proposes a multi-operation mode KPI control method based on regression rule extraction from deep neural networks (R-DeepRED). Firstly, spatio-temporal features of the production process are extracted using slow feature analysis (SFA) and graph attention networks (GAT), followed by classification of operation modes based on the aggregated spatio-temporal features. Subsequently, the proposed method integrates high accuracy of neural network control with high interpretability of rule-based control by employing R-DeepRED method to extract KPI control rules, thereby achieving safe and stable KPI control. The efficacy of the proposed method is demonstrated through a case study of the zinc leaching process, with experimental results validating its effectiveness.