An unsupervised ensemble learning method for real-time anomaly detections in variable refrigerant flow systems

• Published in: Advanced engineering informatics Vol. 68; p. 103648
• Main Authors: He, Haihui, Fan, Cheng, Lei, Yutian, Li, Bin, Chen, Zhijie, Zheng, Chunyuan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2025
• Subjects: Autoencoder; Clustering analysis; Ensemble learning; Fault detection; Variable refrigerant flow system; Clustering analysis; Variable refrigerant flow system; Fault detection; Ensemble learning; Autoencoder
• ISSN: 1474-0346
• DOI: 10.1016/j.aei.2025.103648

•An unsupervised ensemble learning method is proposed for anomaly detection.•The method integrates physics and data characteristics through autoencoders.•A dynamic updating mechanism is developed for online deployment.•The effectiveness of fault detection using various feature groups is compared.•The method achieves a precision rate above 98% in detecting six fault types. Variable refrigerant flow (VRF) systems have gained increasing popularity to accommodate building thermal loads due to its operation flexibilities. To ensure the energy efficiency of VRF systems, it is imperative to establish accurate, automated, and deployable fault detection methods for real-time monitoring and controls. This study proposes an unsupervised data-driven fault detection framework designed for online implementations. Leveraging the concept of clustering and autoencoder ensembles, the reliability of fault detection is improved by considering different input features, clustering algorithms and autoencoder models. The method has been validated using real measurements of 140 VRF machines across various climate zones in China, achieving precision rates higher than 98.6% and recall rates of 78.3–90.2 %. The sensitivity of each feature group under typical faults was compared, providing reference for feature selection. The method developed can be used for identifying suboptimal health conditions for early repairs, reducing energy wastes caused by component or sensor faults, and mitigating negative impacts of operational interruptions.