Data-driven fault detection and diagnosis for packaged rooftop units using statistical machine learning classification methods

This paper proposes and demonstrates a data-driven fault detection and diagnosis strategy for packaged rooftop units using statistical machine learning classification methods. The fault detection and diagnosis task is formulated as a multi-class classification problem. Seven typical rooftop unit faults are discriminated against one another as well as the normal condition. Since experimental data for rooftop units is rare and difficult to obtain, a simulated data library of model faults at steady state operation is used for training and validating the classifications models. Synthetic minority over-sampling technique is used to generate new artificial samples of minority class in order to balance the dataset. Nine well-known classification methods are applied to our dataset, and their performance is compared. The results show that support vector machine, with an overall accuracy rate of 96.2%, is the best classifier, and linear discriminant analysis, with an overall accuracy rate of 76.2%, is the worst classifier. The performance of the classification methods for individual faults is also characterized using true positive rate and false positive rate statistical measures. The relative importance of input variables is also discussed. The high accuracy of the classification methods shows the potential of a data-driven strategy in detecting and diagnosing the rooftop unit faults.