From Experiment to Production: Solution and Evaluation of Centralized Machine Learning System Architecture for Scalable Industrial Predictive Maintenance

Abstract

Industrial predictive maintenance has recently benefited significantly from advances in machine learning. However, most existing approaches remain limited to experimental model level implementations and lack the system-level infrastructure required for reliable deployment in real-world environments. This paper presents the design and empirical evaluation of a centralized artificial intelligence architecture for industrial predictive maintenance, implemented as a scalable, production-grade machine learning system. The proposed approach consolidates data processing, model training, inference, and lifecycle management within a unified backend, while leveraging a distributed microservices infrastructure for execution scalability. The system integrates asynchronous task orchestration, experiment tracking, and model versioning to ensure reproducibility and operational robustness. A comparative study is conducted between an initial synchronous experimental model and the proposed final system. Experimental results on real-world industrial time-series data from a wastewater treatment facility demonstrate significant improvements, the findings highlighting the critical impact of system-level architectural decisions on both machine learning performance and operational scalability. The proposed centralized paradigm provides a practical and reproducible framework for deploying Industry 5.0 guided predictive maintenance solutions in Industrial Internet of Things environments, bridging the gap between academic model initiatives and higher technological readiness level artificial intelligence systems.