SIAD: a low-latency AI-driven software architecture for anomaly detection in industrial IoT environments

In the evolving landscape of the Industrial Internet of Things (IIoT), ensuring the seamless operation of interconnected devices is paramount. Traditional anomaly detection methods often struggle with the volume and complexity of IIoT data, leading to reduced accuracy and increased false-positive rates. To address these challenges, this paper proposes an AI-enhanced software framework termed SIAD (Software for Industrial IoT Anomaly Detection). SIAD integrates lightweight GRU-based inference at the edge and Transformer-based escalation at the cloud to achieve both real-time responsiveness and high detection accuracy. Evaluations using the SWaT and WADI datasets show that SIAD-Hybrid achieves up to 97.4% accuracy, 96.2% precision, 94.6% recall, and 98.1% AUC on SWaT, while maintaining latency below 200 ms. The edge-only variant (SIAD-Edge) achieves response times as low as 87 ms. Compared to baseline LSTM and Transformer models, SIAD significantly reduces false positives by up to 50%. Designed with modular microservices and standard IIoT protocols (MQTT, REST), SIAD ensures scalable deployment and seamless integration into industrial environments. This framework contributes toward reliable, intelligent, and real-time anomaly detection across distributed IIoT systems, with experimental validation in water treatment environments representative of critical infrastructure scenarios.