Real-Time AI and Deep Learning–Driven Data Analytics for Networked Cyber-Physical Systems Using Edge Sensor Hardware

Abstract

Cyber-physical systems being utilized in industrial transportation and critical-infrastructure environments produce heterogeneous sensor streams that require sub-20 ms analytical latency and classification fidelity is not compromised. In this paper, a layered edge-intelligence system that combines temporal convolutional networks with lightweight transformer encoders to deliver real-time anomaly detection and predictive control on edge sensor hardware is introduced. An edge-cloud orchestration protocol is a dynamically partitioned, distributed edge-cloud architecture whose inference is latency-sensitive and that offloads heavy retraining workloads to cloud accelerators. Performances on the Edge-IIoTset dataset, which is further extended to a distributed cyber-physical system design, i.e. 48 heterogeneous sensor nodes connected in a 5G mesh, show an end-to-end inference latency of 17.3 ms, fault-classification accuracy of 96.4%, and an average absolute error of 2.8 cycles to estimate the remaining-use The suggested architecture lowers the bandwidth usage by 61.2% in comparison with cloud-focused baselines and can maintain the performance at the loss rates of packets as high as 18 percent. These findings serve as a benchmark that can be reproduced to enable latency-constrained deep analytics in safety-critical cyber-physical systems.