PHYSICS-GUIDED MACHINE LEARNING FOR SMART SENSING SYSTEMS

Abstract

Machine learning with its superior capability in mapping nonlinear relationships, analyzing complex information, identifying patterns, and learning automatically without being explicitly programmed has achieved tremendous success in many challenging fields and even contributed significantly to novel scientific discoveries. However, the interpretability and generalizability of the data-driven machine learning approach are critical hurdles against wider adoption. This thesis investigates the data-centric aspect of machine learning development and integrates physics-guided solutions to improve the performance of smart sensing systems. Avenues for integrating physics principles with data-centric machine learning are systematically explored and are categorized into five unique approaches addressing the data space, features, quantity, label, and type coherency respectively. Finally, the importance of the above-mentioned approaches is demonstrated using actual smart sensing systems which have not only achieved better model performance, generalizability, and relevancy but also improved the interpretability with consistent physics-based verification.