This paper presents improvements to the statistical rule-based (SRB) fault detection and diagnosis (FDD) technique originally presented by Rossi and Braun. The original method assumed a constant covariance matrix for both normal and faulty operations for fault detection and a diagonal covariance matrix for fault diagnosis, which results in some loss of FDD sensitivity. As a first step in quantifying the loss in sensitivity, Monte-Carlo simulation (MCS) was used to evaluate fault probabilities using a nondiagonal covariance matrix determined from measurements on a rooftop unit. Although MSC provides a good benchmark for off-line comparison, it is not practical for on-line application because of large computational demand. In order to improve the sensitivity of the SRB method without increasing the computation, two new classifiers were developed for the detection and diagnosis steps of the FDD method. The new methods do not require the covariance matrix for faulty operation or probability calculations and are relatively straightforward to implement. In addition, a more robust steady-state detector and an improved modeling scheme are proposed. Finally, a case study is presented that demonstrates the improvement in overall FDD sensitivity as compared with the original method. The data for this comparison were for a rooftop air conditioner that was tested by Breuker and Braun in the laboratory with five different fault types, implemented at five fault levels, and operated in a transient mode at five different load levels.

Units: I-P