A Self-Learning Architecture for Digital Twins with Self-Protection

The digital twin paradigm is a promising enabling technology to accelerate the decarbonisation of industrial sites that use process heat. With digital representations that look-like, behave-like, and connect to a physical system, digital twins bring together critical operational and asset data into a single knowledge store. However, a high-fidelity digital twin relying on the cloud in real-time with direct influence on operations exposes the plant to cyber attacks. We propose a software architecture for a Digital Twin that adaptively generates more accurate representations of its operations to detect malicious activities and mitigate their effects. To achieve this adaptivity, our solution leverages ML, time-series forecasting, concept drift detection and control stability analysis. To evaluate our solution, we develop a simulation of a simple industrial plant consisting of one PID-controlled steam-boiler and a variety of uncertainties. Our experimental evaluation suggests that Dynamic Mode Decomposition with Control, a system identification technique, best contributes towards Self-Learning by producing verifiable models that better align the need for retraining with concept drifts.