The design and viability of a probabilistic fault detection and diagnosis method for vapor compression cycle equipment

The reliable and automated detection and diagnosis of abnormal behavior within vapor compression cycle equipment would be extremely valuable for equipment owners and operators. The objective of this research includes designing and testing the viability of a fault detection and diagnosis system for chillers, which are a specific type of vapor compression cycle equipment. Chillers are commonly the largest, most expensive component within a commercial building's mechanical system. The chiller's purpose is to satisfy building or process cooling loads through the production of chilled water or ice. The author chose the chiller to be the central focus of this research based on several factors. The reasons included the need for improved fault detection and diagnosis systems available on chillers; the large number of chiller installations within buildings around the world; and the availability of a chiller plant within a laboratory setting for experimental research. In addition, chiller faults impose threats to the environment and human health through the emission of ozone depleting refrigerants. Finally, chiller faults typically result in reduced operational efficiency thereby increased energy consumption. The fault detection and diagnostic tool developed here performs the task of analyzing archived chiller operating data and detecting faults through recognizing trends or patterns existing within the data. The author developed a tool to perform these tasks using a neural network. The neural network fault detection and diagnostic program was trained using chiller operating data collected during both normal and abnormal operating conditions. The full scale chiller plant located within the Joint Center for Energy Management Laboratory (JCEM) at The University of Colorado at Boulder was utilized for data collection activities. Faults imposed on the JCEM chiller plant included refrigerant loss and overcharge, oil loss and overcharge, air cooled condenser fouling and the loss of an air cooled condenser fan. Chiller operating data were collected during fault experiments utilizing a realistic load profile. Following the data collection phase, a detailed visualization and analysis was conducted on the archived chiller operating data. The FDD model was trained, validated, and tested using the archived operating data collected in the JCEM Laboratory. The training, validating and testing data sets consisted of data collected during all modes of chiller operation. In addition, each fault mode of chiller operation included several degrees of severity. For example, the refrigerant loss data possessed fault degrees ranging from five percent to sixty percent. This progression occurred in 5% increments over several days of testing. When presented with the validation data, the trained fault detection and diagnostic tool was capable of correctly predicting the current state of chiller operation with a misclassification rate of only 3%. Therefore, the developed fault detection and diagnostic tool for chillers successfully detected abnormal chiller behavior and inferred the most probable cause of the fault.