Lifelong Learning for Monitoring and Adaptation of Data-Based Dynamical Models: A Statistical Process Control Approach

This paper addresses the monitoring and continual learning of data-based dynamical models. Throughout the lifespan of any process, many changes can occur. In an indirect control design framework, in order to maintain an effective control system, it is crucial to monitor the modelling performance and adapt the existing model to possible system variations while preserving previously acquired information. A comprehensive methodology is hence proposed to detect a system-model mismatch and its cause, and to update the model accordingly. The proposed idea consists in leveraging control charts constructed on operational data to spot an anomaly and to determine its cause (endogenous or exogenous). The procedure then provides an adaptation algorithm based on the type of change detected: if endogenous, the model is “partially” updated by means of a Moving Horizon Estimation (MHE) algorithm, if exogenous, the model is “incrementally” updated by means of a model uncertainty estimation algorithm. The proposed methodology is tested in simulation on a district heating system benchmark, showing promising results from the monitoring and continual learning perspective.