Data-Driven Intelligent Systems for Industrial Automation

Abstract

Industrial automation is undergoing a fundamental transition from rule-based programmable logic controller (PLC) systems toward data-driven
intelligent systems capable of adaptive control, predictive maintenance, quality assurance, and autonomous process optimisation — collectively
constituting the operational core of Industry 4.0. Despite widespread deployment of industrial IoT sensors and edge computing infrastructure, the
systematic integration of machine learning into production-grade industrial automation control loops remains constrained by latency requirements,
safety certification demands, and the scarcity of labelled fault data. This study presents IndustrialAI, a unified data-driven intelligent systems
framework for industrial automation integrating a physics-informed neural network (PINN) process model, an online reinforcement learning adaptive
controller, and a semi-supervised anomaly detection module for predictive maintenance. IndustrialAI was deployed and validated at two European
manufacturing facilities: a DTU-affiliated pharmaceutical tablet press line in Lyngby (48-sensor continuous manufacturing, 3.2M tablets/day) and a
University of Florence-partnered automotive body welding line in Florence (72-sensor, 840 weld joints/hour). IndustrialAI achieved process yield
improvement of 4.7%, energy consumption reduction of 11.3%, predictive maintenance recall of 0.963 (7-day fault horizon), and false alarm rate of
1.4% — improvements of 28.4%, 17.6%, and 61.8% over incumbent PLC-based and threshold systems respectively. Control loop adaptation latency
averaged 47 ms, within industrial real-time requirements. These results demonstrate that physics-informed, safety-aware data-driven intelligent
systems are technically deployable and operationally beneficial in real European manufacturing environments.