Conference Paper
Proceedings of SimBuild Conference 2024
Adaptive Fault Detection and Diagnosis Based on Growing Gaussian Mixture Regressions for Passive Chilled Beams System
Adaptive Fault Detection and Diagnosis Based on Growing Gaussian Mixture Regressions for Passive Chilled Beams System
Sujit Dahal 1, Liping Wang 1, James E. Braun 2,31 Civil and Architectural Engineering, University of Wyoming, USA 2 School of Mechanical Engineering, Purdue University, USA 3 Center for High Performance Buildings, Ray W. Herrick Laboratories, Purdue University, USAAbstract: A novel learning-based fault detection and diagnosis (FDD) strategy utilizing a Gaussian mixture model (GMR) has been proposed to overcome the limitations of conventional FDD methods, which struggle with unidentified faults. This approach adapts to incorporate information about unknown faults. The method was tested using data from a passive chilled beams (PCB) system, an area with limited FDD research. The model was initially trained with winter data and later evaluated with spring and summer data. Faults specific to PCB systems were identified from literature and practical experience and simulated using building automation software. A feature selection method was employed to distinguish between normal and faulty operations. The GGMR FDD model was developed using winter data for known faults and normal operations. When encountering new fault data classified as unknown, the model evolves by updating and adding Gaussians. This process enables the incorporation of the new fault information. The model's performance was initially tested with winter data and then across different seasons. While the static model was less effective during the spring and summer due to changes in system features, the evolved model demonstrated high efficacy in adapting to new fault patterns and ensuring accurate fault diagnosis. The evolved model achieved over 92% fault prediction accuracy for spring and summer, highlighting its adaptability and effectiveness in varied conditions. Keywords: Evolving learning, Fault detection and diagnosis, Feature selection, Passive chilled beams, Seasonal comparisonPages: 818 - 829 Paper: