Edge AI-driven adaptive maintenance for manufacturing systems: improving decision-making and reducing energy consumption

Purpose This study develops and validates an Edge AI-driven adaptive maintenance framework aimed at improving decision-making and reducing energy consumption in energy-intensive manufacturing systems. Design/methodology/approach Using a real-world case study from Qatar Steel Company, we collected multimodal sensor data (vibration, acoustic and thermal) and integrated advanced preprocessing techniques, including Fast Fourier Transform (FFT) for vibration, noise reduction for acoustics and smoothing for thermal signals. A sensor adaptation module (SAM) was introduced to harmonize heterogeneous features before feeding them into a Random Forest-based Edge AI model. Additionally, PCA-based feature compression and adaptive edge resource allocation were implemented to optimize computational performance and energy efficiency. Findings The proposed framework improved predictive accuracy, achieving an AUC of 0.91 while reducing data size by 68% through feature compression and decreasing edge device CPU utilization by 27% via dynamic resource scheduling. These results demonstrate the framework’s capability to deliver energy-aware, real-time adaptive maintenance in practical industrial environments. Research limitations/implications The framework provides a ready-to-deploy decision-support tool for energy-intensive manufacturing sectors, reducing downtime, improving energy efficiency and enabling broader implementation of Industry 4.0 strategies. Practical implications The model offers a practical decision-support tool for maintenance engineers, enabling proactive intervention, minimizing downtime and aligning maintenance with energy-saving goals in smart factory settings. Originality/value This study introduces a novel and practical Edge AI-driven adaptive maintenance framework that integrates multimodal sensor fusion, feature compression and adaptive resource allocation to enhance energy efficiency in manufacturing systems. Unlike previous studies, we developed a sensor adaptation module (SAM) to unify heterogeneous sensor data (vibration, acoustic and thermal) into a consistent feature space, enabling robust predictive performance across diverse operating conditions. Furthermore, the incorporation of PCA-based feature compression and dynamic edge resource optimization distinguishes our work from prior cloud-based or simulation-focused approaches. These contributions provide a deployable solution tailored for energy-intensive environments and advance the practical adoption of Edge AI in smart manufacturing.