Predictive maintenance of HVAC system using supervised machine learning algorithms

Abstract

Technological advancements, especially rich data, being incorporated into the development of buildings is simultaneously elevating the importance of fault detection and diagnostics (FDD) of energy-e cient applications (for instance, an HVAC system). Consequently, the challenges lie in identifying the suitable FDD techniques and test methodologies and locating the appropriate dataset. Furthermore, identi cation and detection of fault during early stages and later stages determine how critical the problem is and what form of immediate maintenance is required (i.e., replacement of components, altering a given condition to restore normalcy, also known as negative feedback etc). This is especially important in shopping malls, educational and business institutions, where the malfunctioning of air conditioners or dehumidi ers can greatly a ect productivity. The prediction models revolve around the maintenance of condition-based de nitions for the ground truth of the system. At rst, the preprocessing technique of normalising the data set was performed on it. Then the data was visualised to help identify categorical and important features in uencing the binary values of the fault detection ground truth. Afterwards, the dataset was split into train-test sets with di erent machine learning algorithms, namely RandomForestClassi er, Decision Tree, KNearestNeighbours, Catboost, MLP Classi er etc. From our experiment, we achieved the best testing accuracies of 99.21% and 100% using the Random Forest classi er on the MZVAV-1 and SZCAV datasets respectively, 95.76% using the CatBoost classi er on the MZVAV-2-2 dataset, 99.91%, 100% and 99.71% using the Extra Trees classi er on the MZVAV-2-1 dataset, SZCAV and SZVAV datasets respectively.