Abstract
The Industrial Internet of Things (IIoT) generates massive sensor data streams that require real-time anomaly detection to prevent equipment failures and ensure operational safety. Traditional cloud-centric approaches suffer from high latency and bandwidth constraints, motivating the adoption of edge AI. This paper proposes a hybrid edge AI framework that combines lightweight machine learning models with rule-based reasoning for real-time anomaly detection in IIoT environments. The framework deploys a quantized convolutional neural network (CNN) and a rule engine on edge devices, with a cloud backend for model updates and retraining. Using a public multivariate time-series dataset from a water treatment plant, we evaluate detection accuracy, latency, and resource utilization. Results show that the edge-deployed model achieves 96.2% detection accuracy with an average inference latency of 12.3 ms, reducing data transmission to the cloud by 85%. Comparative analysis with cloud-only and baseline edge methods demonstrates significant improvements in real-time performance and bandwidth efficiency. The findings highlight the viability of edge AI for industrial anomaly detection, with implications for predictive maintenance and operational resilience.