Fault Classification From Multivariate Time Series in Fan Coil Units: The Impact of the Temporal Component

The rapid detection and accurate identification of faults in building systems are essential for minimizing the effects of malfunctions and optimizing resource efficiency. However, fault diagnosis remains a complex task, particularly when it involves the classification of multivariate time series data. In this study, we investigate how temporal dynamics influence the performance of both standard and multivariate time series classifiers, with a focus on fault detection in fan coil units. Additionally, we evaluate the complexity of different fault identification tasks in this context, using classifier performance as a key metric. We further evaluate the classifiers' generalization capabilities by examining their accuracy as the temporal gap between the training and testing datasets is widened. Our findings highlight several variants of random forest classifiers and the gradient boosting classifier as the top-performing methods for this problem.