Dynamic system identification through inverse optimization models for industrial process monitoring

Abstract

This paper proposes a novel dynamic system identification approach based on inverse optimization for the monitoring of industrial processes. The proposed method recovers unknown parameters of an objective function, trade-offs, and process decision-making mechanisms from time-indexed data gathered via industrial sensors, SCADA systems, energy meters, and production logs. In contrast to traditional forward identification approaches that only identify input-output relationships, the proposed framework aims at reconstructing the optimization procedures behind observed process decisions. An optimization model with a parametric objective function, observables of the operational context, decision variables, and feature mappings, along with engineering constraints, is proposed. Unknown parameters are recovered via minimizing inconsistencies of the Karush-Kuhn-Tucker optimality conditions in multiple decision periods, which allows us to account for time evolving process conditions and near-optimal behavior. The scale normalization is used to derive the trade-off ratios for the operational goals that can be interpreted. The derived dynamic model helps in the process of monitoring the process continuously, predicting its trajectory, detecting any parameter drifts, and determining abnormal operation modes. Rolling window estimate allows adapting the model to possible structural changes in the process behavior. The approach ensures the trade-off between predictive quality, constraints and interpretability, overcoming the drawbacks of the traditional black box identification models. Thus, in general, the work creates an inverse optimization-based framework for translating decisions made in real-world industry into dynamic and interpretable process models.