JoWUA

Abstract

The integration of Edge Artificial Intelligence (Edge AI) with the Industrial Internet of Things (IIoT) is enabling real-time decision-making in smart industrial environments by reducing latency, bandwidth usage, and reliance on centralized cloud systems. Nevertheless, there is still a challenge of choosing an ideal Edge AI architecture, because of performance, energy efficiency and scalability trade-offs. In this paper, the complete performance analysis of edge-native, hybrid, and cloud-assisted artificial intelligence systems is given regarding the performance of such systems in the industrial domain in predictive maintenance, anomaly detection, and visual inspection. An experimental heterogeneous environment was created with the help of various edge devices, such as GPU-enabled and TPU-based systems, and lightweight deep learning models, such as MobileNet, YOLOv5n, and LSTM. Industrial sensor, visual, and acoustic data benchmark datasets have been used. The main performance indicators such as inference latency, energy usage, and model accuracy and network bandwidth were tested in different work load and environmental conditions. It has been experimentally demonstrated that edge-native architecture can offer ultra-low latency (720 ms) which is much lower than cloud-based systems (100-300 ms). The platforms that were energy efficient showed consumption as low as 0.3 J/inference and had model accuracy of up to 95.6 in visual inspection conditions. Also, the predictive maintenance use cases reported up to 81 percent in equipment downtime. The hybrid architectures enhanced accuracy by trading off edge and cloud intelligence but came with the moderate latency overhead (~80 ms). The results point at the fact that latency-sensitive applications are best served with edge-native deployments whereas hybrid solutions offer a tradeoff between performance and scalability. This study offers practical design insights for deploying efficient and reliable Edge AI systems in industrial environments and outlines future directions in federated learning and adaptive edge intelligence.