Context Aware Model Learning in Cyber Physical Systems

The operational behavior of Cyber-Physical Systems (CPS) can vary in real-world settings due to unpredictable elements of the dynamic environment or human interactions. It is crucial to identify these changes promptly to ensure the system operates efficiently and safely. One approach is online model learning, which continuously learns an operational model to spot any discrepancies. Existing model learning techniques, while effective, often require substantial data inputs and incur high computational costs making it unsuitable for real-time safety monitors. In this paper, we introduce a methodology that starts with learning the model offline, considering the context, and then simplifies the online model learning challenge into linear or polynomial regressions through the Data Context Driven Model Reduction (DCDMR) strategy. DCDMR exploration of the operation of a CPS uncovers the underlying changes in the model characteristics. We demonstrate the capability of the proposed technique for learning the overall model of the Artificial Pancreas controller.