Unsupervised anomaly detection for HVAC units

This thesis project focuses on the detection, using unsupervised methods, of anomalies in HVAC units. The main goal is to detect compressors' faults by training a specialized isolation forest model on raw sensors' data. To this end we conducted several experiments tuning the hyperparameters of the model, studying the results using a depth-based feature importance tool for the explainable artificial intelligence task and asking feedbacks to the domain experts. During this process, we managed to reduce the features from over 200 to nearly 30 and improve the accuracy of anomaly detection for compressors' faults.