A Model-Based Fault Detection and Diagnosis Methodology for HVAC Subsystems

Abstract

The existence of faults in Heating, Ventilating, Air-Conditioning and Refrigerating (HVAC&R) systems plays a significant role in the degradation of comfort levels for building occupants on one scale, and the degradation of quality of life on earth, on a larger scale. The objective of this research work is to develop a robust Model-Based Fault Detection and Diagnosis (FDD) Methodology for a chiller subsystem that is expandable, transportable and suitable for online implementation. It shall be expandable in the sense of being able to accommodate more complicated phenomena and transportable in the sense of being able to model different systems and various refrigerants.

Such an FDD methodology is eventually attained by first, modeling the chiller subsystem using a combination of energy balances, material balances, fluid properties and equipment geometry to establish physical relations between all variables involved (both dependent and independent). That is, a physical model is developed.

Next, computer simulation techniques are used to investigate the performance of the subsystem within prescribed limits of the input variables. Calculated characteristic quantities (CQ) in the model are used to determine various fault conditions. A General Regression Neural Network is also trained and calibrated to learn the base-case (fault-free) behavior of the chiller.

Lastly, the FDD methodology for detecting and diagnosing faults is designed to compare actual operating conditions to the corresponding base-case conditions, compute CQ residual errors and use the appropriate statistical analyses to determine their overall significance. If these residuals are significantly different from zero, a fault detection alarm may be justified. Otherwise, silent monitoring continues with the assumption that the chiller process is operating normally and that no faults are present.

Robustness of the methodology is obtained by accounting for the accuracy of the sensors used in obtaining measurement data. This is done by first determining the magnitude of errors in sensor measurements and then using them to conduct a general uncertainty analysis to examine how they propagate into the CQs that are computed using these sensor measurements. Fault Detection thresholds are subsequently derived and alarms are made more confidently when these thresholds are exceeded.