Bayesian method for HVAC plant sensor fault detection and diagnosis

Together with the thermo-physical relationships among the flow rates and temperatures of water in a piping system, the Bayesian method was employed to develop a model for detection and evaluation of biases of water flow and temperature sensors in a central chiller plant. The model can handle biases of multiple sensors occurring simultaneously and can remain functional when the coverage of the available measurements is incomplete. A series of case studies was done to verify the performance of the model and for comparison with the conventional method that is based solely on the thermo-physical relationships. The cases studied involved the use of synthetic plant operating data and actual operating records of an existing chiller plant. In this paper, the theoretical basis of the model is outlined, and explanations are given for the superior performance of the Bayesian method in handling cases with data that cannot fully cover the required range of operating chiller patterns. Results of the cases unveiled the effects of the prior belief with or without being updated during the estimation process, and of biases occurring in steps at the same time and at different times, as well as those that would increase with time. Furthermore, the case studies showed that the Bayesian method was able to detect sensor biases of a magnitude of +/- 0. 5 degrees C or lower. (C) 2020 Elsevier B.V. All rights reserved.