Buildings constitute 36% of the world’s energy consumption of which up to 30% could be wasted due to improper controls tuning, inadequate maintenance and degradation of equipment. Many of these faults are never detected or at least difficult to detect. Hence, Automated Fault Detection and Diagnosis (AFDD) tools are needed to detect the presence of these faults in the system. The main objective of this research is to perform a comparison of the performance of supervised learning-based Machine Learning (ML) algorithms for time series-based fault detection in the Air Handling Unit (AHU). Additionally, the performance of algorithms with a lower feature set (without any symptomatic feature) was compared to the performance with a higher feature set (possibly including symptomatic features). A real office building (Breda, The Netherlands) served as a living lab where several low-cost monitors and mass flow rate meters were installed and data was logged at a 1-min interval by the Building Management System (BMS) to obtain high-granularity data. Non-faulty, baseline data was used to train models based on seven algorithms: Support Vector Regression (SVR), Random Forest (RF), Decision Tree (DT), XGBoost, CatBoost, LightGBM and Multi-Layer Perceptron (MLP). The best performing algorithms were evaluated for application to symptom detection on data collected through fault experiments. The performances of the algorithms on symptom detection were evaluated using symptom detection metrics. The results showed that the ML models performed better with a greater number of features and XGBoost performed the best in majority of the tested scenarios.