Non-Linear Recursive Parameter Estimation Applied to Fault Detection and Diagnosis in Real Buildings

This paper describes a fault detection and diagnosis scheme based on parameter estimation and presents results from its application to a cooling coil subsystem in a real building. A non-linear adaptation of the Prediction Error Forgetting (PEF), algorithm is employed to estimate the parameters of a simple, steady-state, first principles based cooling coil model. A steady-state detector is used to discard data with excessive transients. Three model parameters represent possible faults; control valve leakage, coil fouling and sensor offset. These parameters are estimated recursively, together with the uncertainty in the estimated values. A significant change in a particular parameter indicates abnormal operation and suggests a diagnosis. The paper describes the first-principle models and their fault parameters, the steady-state detector, and the recursive parameter estimation algorithm. Results from the application of the technique to data measured in a test building demonstrate that valve leakage and coil fouling can be detected and diagnosed. The applicability of the approach to fault detection and diagnosis in real systems is also discussed.