Adaptive learning based seasonal fault detection and diagnosis of radiant heating and cooling system using Gaussian mixture regression

This study presents an adaptive Gaussian Mixture Regression (GMR) framework for the fault detection and diagnosis (FDD) of radiant slab (RS) heating and cooling systems. This domain has not been extensively explored in FDD research. Our study addresses the challenge of identifying and adapting to previously unknown or unexplored operational states within HVAC systems, exacerbated by the influence of seasonal changes on fault characteristics. An evolving GMR model is developed that learns from the dynamic performance of building systems and their components across different seasons. Utilizing a comprehensive dataset collected from a real office environment during winter, spring, and summer, the model is rigorously tested against various faults. A feature selection process is applied to determine the most indicative features for fault diagnosis. The adaptive mechanism allows for updating the existing or integrating new Gaussian components upon detecting unknown states, thereby preserving its diagnostic precision over time. The GMR FDD model was trained with winter data and tested with winter, spring, and summer data. Our findings reveal that fault diagnosis accuracy is extremely high for winter, both before and after evolving the model, but is compromised without the model evolution for the data from the spring and summer seasons. However, once evolved, the model demonstrates a remarkable ability to assimilate new fault patterns, achieving high diagnostic accuracy rates—99.6% in spring and 97.3% in summer—thereby confirming its robustness and adaptability to seasonal variations in fault diagnosis.