Data-Driven Fault Detection and Diagnosis in Cooling Units Using Sensor-Based Machine Learning Classification

Precision air conditioning (PAC) systems are prone to various types of failures, leading to inefficiencies, increased energy consumption, and possible reductions in equipment performance. This study proposes an automatic real-time fault detection and diagnosis system. It classifies events as either faulty or normal by analyzing key status signals such as pressure, temperature, current, and voltage. This research is based on data-driven models and machine learning, where a specific strategy is proposed for five types of system failures. The work was carried out on a Rittal PAC, model SK3328.500 (cooling unit), installing capacitive pressure sensors, Hall effect current sensors, electromagnetic induction voltage sensors, infrared temperature sensors, and thermocouple-type sensors. For the implementation of the system, a dataset of PAC status signals was obtained, initially consisting of 31,057 samples after a preprocessing step using the Random Under-Sampler (RUS) module. A database with 20,000 samples was obtained, which includes normal and failed operating events generated in the PAC. The selection of the models is based on accuracy criteria, evaluated by testing in both offline (database) and real-time conditions. The Support Vector Machine (SVM) model achieved 93%, Decision Tree (DT) 93%, Gradient Boosting (GB) 91%, K-Nearest Neighbors (KNN) 83%, and Naive Bayes (NB) 77%, while the Random Forest (RF) model stood out, having an accuracy of 96% in deferred tests and 95.28% in real-time. Finally, a validation test was performed with the best-selected model in real time, simulating a real environment for the PAC system, achieving an accuracy rate of 93.49%.