Approche basée sur le transformateur pour la détection des défaillances de systèmes d’eau réfrigérée : éclairages à partir des domaines temporels et fréquentiels.

The maintenance of heating, ventilation, and air conditioning (HVAC) equipment is a critical research area, particularly for anomaly detection and diagnosis in chilled water systems. While numerous methods have been proposed to achieve satisfactory detection and false alarm rates, their effectiveness in real-world environments remains limited due to challenges such as difficulty in detecting unseen anomalies, reliance on large amounts of anomalous data, and lack of adaptability to different parameter settings. To address these issues, this study focuses on identifying anomalous data by analyzing the normal operations of chilled water systems and examining their characteristics in the frequency domain over both short- and long-term periods. Based on this analysis, two distinct anomaly detection schemes are proposed: a low-frequency detection (LFD) method, which focuses on low-frequency data, and a frequency-domain segmentation detection (FDSD) method, which separately considers high- and low-frequency components. Unlike previous approaches, the proposed LFD and FDSD comprehensively integrate both time-domain and frequency-domain characteristics while determining anomaly detection thresholds without requiring anomalous data. Experimental evaluations on the ASHRAE RP-1043 dataset demonstrate that FDSD achieves a detection rate of over 90 % under varying background conditions for subtle anomalies that are typically difficult to detect. Compared to FDSD, LFD achieves a satisfactory detection rate of over 80 % while offering advantages such as shorter training time and lower model complexity.