Abstract

Fault detection and diagnosis (FDD) systems for Heating, Ventilation, and Air Conditioning (HVAC) operate reactively, identifying failures only after they occur—leaving building operators unable to prevent equipment failures, occupant health consequences, and system failures. This study establishes a sequential deep learning framework demonstrating that multi-step-ahead forecasting fundamentally outperforms single-point classification for early fault detection. Through controlled filter degradation experiments across four progression stages (normal, incipient, early-warning, critical-warning) in multi-unit HVAC systems, we employ capacity-agnostic workload normalization to enable learning across heterogeneous unit capacities. Systematic Bayesian optimization across four sequential architectures (Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM)) reveals three critical discoveries. First, LSTM with extended temporal memory windows optimally captures progressive degradation patterns, demonstrating that gated mechanisms are essential for learning fault evolution over time. Second, reconfiguring from single-point classification to multi-step-ahead forecasting dramatically improves early-stage detection (F1-score : 0.44 to 0.83) while reducing false alarms by 2.6-fold (Precision : 0.34 to 0.89). The key mechanism: faults masked during low-demand periods manifest clearly under high-demand conditions. Third, performance analysis establishes a critical operational threshold: early-stage faults are perfectly detected (F1-score : 1.00) when the system workload exceeds baseline demand, demonstrating that detectability depends on operational intensity rather than fault age. This framework transforms fault detection from reactive diagnosis to predictive monitoring, preventing critical failures before they escalate in multi-unit HVAC systems.