Comparative analysis of four classification algorithms for fault detection of heat pumps

Abstract

Heat Pumps (HPs) technology had a remarkable diffusion and growth in the Heating, Ventilation, and Air-Conditioning (HVAC) field, with an increase predicted to be greater than 200% until 2030. Therefore, efficient asset management must be achieved by ensuring their maintainability and reliability. For this purpose, several methodologies have been used in the last years for HPs fault diagnosis. These methodologies include, with the advances of computational systems and reduced costs of several components, machine learning and other artificial intelligence techniques. In the present paper, a connection weight approach is used to identify the clout of each component for the fault detection (FD) of HPs. The results show that the method provides a more comprehensive assessment of the importance of each component, yielding a reduction of 50% of the features. A comparative analysis of supervised learning classification algorithms is performed, for fault detection of an air-to-air HP in cooling mode. Such algorithms include Naïve Bayes, Support Vector Machine, Logistic Regression, and K-nearest Neighbors. The faults considered include compressor and reversing valve leakage, improper condenser and evaporator fouling, liquid line restriction, refrigerant undercharge and overcharge, and also the presence of non-condensable gases. Algorithms are compared in terms of accuracy, precision, recall, and F1 Score. The best results were achieved with K-nearest Neighbors, since it presented the highest values for these four metrics, over 99% after a cross validation considering a reduced number of components. Naïve Bayes, Support Vector Machine, and Logistic Regression also achieved over 90% performance metrics. Other models were superficially analysed and Ridge Regression with metrics performance of 100%, on average, revealed to be a promising tool for fault diagnosis in HPs.