Data-Driven Smart Sensing and Real-Time High-Fidelity Digital Twin for Building Management

Abstract

Achieving a high-fidelity digital twin of buildings is desirable but often requires a substantial influx of real-time data, necessitating a dense network of environmental sensors. In addition, high hardware expenses, deployment costs, constraints, and sensor malfunctions can impede the realization of such digital twins. In this article, we introduce TwinSense, a pioneering system that harnesses data-driven virtual sensing within a real-time high-fidelity 3-D digital twin of the building environment. Our innovative method utilizes machine learning (ML)-based inference to accurately estimate real-time sensor variables, such as temperature and $\text{CO}_{2}$ levels, across both 2-D (varying rooms) and 3-D (diverse elevations) domains, with limited reliance on physical sensors. By extending sensing coverage through ML-driven virtual smart sensing, we create a more accurate digital twin for advanced building management. Our case study results across different seasons indicate an average mean absolute percentage error of 2% for temperature and approximately 5% for air quality parameters, such as $\text{CO}_{2}$, when compared against ground truth physical sensors deployed in the multiroom segmented building. Furthermore, we propose a modified 3D-inverse distance weighting thermal interpolation method. Leveraging the comprehensive multicamera-angle visualization capabilities facilitated by the Unreal Engine, our analysis revealed a potential anomaly within the air conditioning system. The building management team validated this observation during the real-world trial, affirming the initial efficacy of our solution.