Machine Learning-Based Anomaly Detection and Classification in 36,000 BTU Air Conditioners

This research proposes a novel approach for diagnosing and analyzing faults in air conditioning (AC) systems using artificial intelligence (AI) to classify failure types. The study focuses on developing a fault diagnosis system for a 36,000 BTU air conditioner to reduce human decision-making errors and improve diagnostic efficiency. Machine learning techniques are used to classify system anomaly faults into 5 categories: (1) Normal Operation, (2) Dirty Indoor Unit, (3) Low Refrigerant Level, (4) Insufficient Outdoor Heat Dissipation, and (5) Others. Operational data were recorded using a Transmission Monitor (TM2016), comprising 54 features including temperature, current, and control signals under both normal and abnormal conditions. Due to the small and imbalanced dataset, the Synthetic Minority Over-sampling Technique (SMOTE) was applied during preprocessing. The dataset was then split into training and test sets. A 5-fold cross-validation was used for model evaluation. Among the evaluated models, the Random Forest demonstrated the highest prediction performance with an accuracy of 88.75% on the training data and 44.44% on unseen data. The Random Forest successfully diagnosed a “low refrigerant” fault, which was later confirmed and repaired by experts. After repair, the model accurately identified normal operation, aligning with expert assessments. The proposed machine learning approach significantly enhances the efficiency of AC fault diagnosis. While expert technicians typically spend hours manually analyzing data, the developed model delivers accurate diagnoses within minutes, enabling rapid evaluation of numerous AC units. This not only reduces the workload for technical personnel but also improves diagnostic accuracy and allows technicians to focus on novel or uncertain cases.