On-Line Models for Use in Automated Fault Detection and Diagnosis for HVAC&R Equipment

To improve overall fault detection and diagnostic (FDD) performance, many of the techniques proposed for automated FDD rely on the use of steady-state models for expected values of operating states under normal operating conditions. Furthermore, on-line measurements and models for overall performance (e.g., EER) are useful in evaluating whether faults are serious enough for service to be performed. This paper presents a modeling approach for operating temperatures that combines a low-order polynomial with a general regression neural network (GRNN) to achieve both good interpolating and extrapolating performance. Using experimental data for a rooftop air conditioner, the performance of this modeling approach was evaluated and compared with several other black-box modeling approaches, including simple polynomials, pure general regression neural networks (GRNN), back-propagation neural networks, and radial-basis functions. This approach was further validated using data collected from a field site at Purdue University. In addition, a simple method for estimating cooling capacity, power consumption, and EER using low-cost measurements is presented and evaluated.