Where Fault Detection and Diagnosis Meets MPC Performance Assessment: Review and Case Study of an Integrated Framework

Various methodologies have been developed over the years to assess model predictive control (MPC) performance. However, few have been applied in industry, and they remain limited in terms of providing a rapid indication of the root causes of deteriorated control performance. This article aims to review the existing methodologies in the literature that address these challenges. Additionally, it implements a structure in which MPC performance assessment is integrated within a fault detection and diagnosis (FDD) framework. The integrated approach employs cascaded modules of machine learning (ML) binary classifiers arranged in a sequence that mimics the decision-making logic of an operator. To illustrate the integrated strategy both conceptually and operationally, a van de Vusse reactor, controlled by a nonlinear model predictive controller (NMPC), is used as a case study. The ML models evaluated include XGBoost, Random Forest, Multilayer Perceptron, Long Short-Term Memory, and Gated Recurrent Unit. The results show that these models can correctly distinguish the cause of abnormalities, even in the presence of measurement noise, with a detection accuracy of 99% and an abnormality classification accuracy above 85% for the best-performing models. Different ML models performed best for distinct diagnostic tasks, highlighting the flexibility of arranging models according to their most suitable application. The investigation indicates that the proposed ML-based FDD framework, which embeds control performance assessment, is competitive for control-aware diagnosis of MPC-controlled processes.