Comparative investigation of seven supervised regression reference models for chiller fault detection: generalization to unseen days and detection at a matched false alarm rate

Abstract Residual-based chiller fault detection and diagnosis (FDD) relies on a fault-free reference model, but reference models are usually compared on data from the training day and without statistical tests. This study compares seven supervised regression methods implemented in MATLAB (fractional logistic regression, lasso, elastic net, decision tree, random forest, least squares support vector regression and a multilayer perceptron) as reference models for six feature parameters of an ASHRAE RP-1043 centrifugal chiller. Accuracy is evaluated on the training day, on two held-out normal days and by leave-one-day-out prediction with Diebold–Mariano and Friedman–Nemenyi tests; detection is evaluated on 26 fault cases at deployable and matched false alarm rates. Lasso and elastic net gave the lowest unseen-day errors for five of the six feature parameters. With thresholds set from training data, the linear models kept the unseen-day false alarm rate at 1.2–2.7%, whereas the tree-based models reached 6.7–9.2%. At a matched 1% false alarm rate, six models gave similar mean detection rates (33.7–35.7%), and least squares support vector regression was markedly lower. Most of the residual variance on normal days consisted of offsets persisting over whole operating conditions, which bounds detectability. Condenser fouling became detectable only with 45% of the tubes plugged, because lower levels reduced the condenser UA by only 5–13%, within the fault-free scatter. A sensor ablation showed that the condensing pressure and power measurements carry the refrigerant-side and water-side faults, respectively. The reference model mainly determines how reliably a threshold transfers to new days.