Integration of digital twins and machine learning for predictive maintenance using APAR method rules in non-residential buildings

Introduction Non-residential buildings are major global energy consumers, and HVAC efficiency is therefore critical. Air Handling Units are central to indoor climate control but are prone to operational faults due to complex control dynamics. Current fault detection approaches often suffer from limited interpretability in machine learning models and limited flexibility in purely rule-based methods. Methods We developed a hybrid predictive maintenance framework that combines interpretable fault detection using Air-Handling Unit Performance Assessment Rules, adaptive fault classification and prediction using machine learning, and near-real-time monitoring through a digital twin interface. The framework was deployed on an AHU in a non-residential facility in Grimstad, Norway, using six months of operational data with more than 51,000 logged records. Results The hybrid approach improved fault detection performance across both frequent and rare fault classes, achieving strong F1-scores and high recall for critical fault conditions. The digital twin component, integrated through pyRevit and a web-based dashboard, enabled near-real-time fault visualization and supported maintenance planning. Discussion The results indicate that combining expert-driven rules with machine learning and digital twin technology can deliver a practical, accurate, and scalable solution for predictive maintenance of AHUs. The framework supports the transition from reactive to intelligent building operations and can be adapted to similar non-residential contexts.