Adapting deep learning-based sensing systems to cyber-physical dynamics
Jiale Chen · 2025
In cyber-physical systems (CPS), the computational and physical elements are deeply coupled. Sensing is an important aspect of CPS. Recent CPS sensing features the use of deep neural networks (DNNs) that often deliver state-of-the-art accuracy. However, the deployment of DNNs must consider the changeable nature of CPS stemming from the dynamics of the cyber and physical elements, which can be highly unpredictable depending on the application characteristics. They impose signi cant challenges in meeting the performance requirements of CPS sensing, such as accuracy, timeliness, and stability. Existing studies have proposed various approaches to adapt DNN-based sensing to cyber-physical dynamics, including model compression, dynamic model deployment, and model scheduling. However, these approaches often su er from either accuracy loss or failure to meet real-time requirements. This thesis aims to design approaches for adapting DNNbased sensing to the cyber-physical dynamics in maintaining sensing performance. Speci cally, it aims to identify the system con guration and behavior that can maximize performance while meeting all the constraints based on the observation of cyber-physical dynamics. The problems to be addressed and the proposed approaches are presented in three studies forming the main technical contributions of this thesis.