Edge-AI Architecture for Real-Time Cyber-Physical System Control Using Embedded Controllers and Cloud Data Analytics

Abstract

The integration of embedded computing, wireless networking and cloud-scale data analytics has revolutionized how industrial and autonomous cyber-physical systems (CPS) are monitored and controlled. Currently, responsiveness, energy efficiency, and resilience to communication disruptions must be part of any competitive CPS. Conventional centralised control architectures, which transmit all sensed data to remote servers before issuing actuator commands, introduce structural lapses incompatible with the time-critical loops of processes in manufacturing automation, smart grid, autonomous vehicles and precision agriculture. A new architecture in which lightweight inference and closed-loop control run directly on embedded microcontrollers mitigates these weaknesses while reducing upstream bandwidth consumption and increasing operational redundancy. This paper proposes a layered Edge-AI architecture that enables CPS control in real time while integrating sensor acquisition and on-device signal conditioning, followed by the deployment of two competing control models, the proportional-integral-derivative (PID) and fuzzy logic, on the embedded controller node. Subsequently, telemetry forwarding is structured to support the cloud analytics backend. The performance of these two models is compared and contrasted using integral squared error (ISE), integral absolute error (IAE), root-mean-square error (RMSE) and overshoot indices. According to the comparative results, the fuzzy logic controller achieves an ISE that is 21% lower than the PID controller's and an RMSE that is 18% lower, when both controllers are subjected to a step disturbance. Both controllers maintain their local control-loop latencies below 5 ms. Consequently, this is well below the sub-10 ms budgetary requirement of Class-A CPSs. The designed architecture also includes OTA (Over-the-Air) model update pathways, enabling continuous edge inference without blocking. The results offer actionable design guidance for engineers deploying intelligent edge controllers in CPS settings with limited resources and tight latency constraints.