Enhancing automated fault detection in building systems: A percentile-based scoring approach with dynamic tolerance range for residual evaluation

Fault detection and diagnostics (FDD) in heating, ventilation, and air conditioning (HVAC) systems using machine learning (ML) methods within a residual-generating approach is a promising solution to overcome obstacles in practical
application. This paper proposes a residual evaluation method with a dynamic tolerance band and residual scoring for FDD processes. The score functions are automatically determined using the percentiles of the residual distribution during the training phase. The transferability of the method is demonstrated using datasets from nine different air handling units, and the performance of fault detection (FD) is evaluated based on a threshold for the total residual score. Compared to a static method based on the L2 norm, the proposed method significantly reduces
the number of false alarms, which is crucial for its acceptance in practical applications.