Beyond Raw Data: Physics-based Dimensionless Features for Fault Detection in Thermal Systems

This thesis presents a novel approach to the fault detection problem in HVAC/R systems, based on dimensionless quantities named Pi Groups. The method relies on the identification, validation and test of suitable Pi Groups with the aim of employing them as features for a more straightforward, intuitive and efficient application in detection algorithms for anomaly identification. This work follows the development of a pipeline responsible for each step: firstly Pi Groups are selected and their choice justified, then their aptitude for the task at hand is validated thoroughly through statistical and heuristic methods, lastly their performance is tested. The workflow is then evaluated by applying it on a realistic heat pump model developed in Simscape. In the last part of this work, the advantages of this novel method are discussed and comparisons are drawn between Pi Groups and the standard procedure based on raw variables, and further developments are suggested.