The rapid adoption of IoT-based smart-building technologies is enabling predictive maintenance that outperforms traditional reactive or purely scheduled strategies. This work presents a practical end-to-end workflow for predictive AHU maintenance by combining real-time sensor data, machine learning, and a BIM-based digital twin. In a case study, an industrial AHU system was created with IoT sensors (18 zone temperatures) and simulated over 11 days under both normal conditions and injected faults (e.g. valve malfunctions, stuck valves, filter blockages, fan stress), yielding labeled datasets of healthy vs. faulty behavior. After data cleaning and feature engineering to capture temporal patterns, two complementary models were trained: a Random Forest classifier for real-time fault detection and diagnosis, and a Long Short-Term Memory (LSTM) network for sequence-based early fault prediction.