Demonstration of Fault Detection and Diagnosis Methods for Air-Handling Units

Results are presented from controlled field tests of two methods for detecting and diagnosing faults in HVAC equipment. The tests were conducted in a unique research building that featured two air-handling units serving matched sets of unoccupied rooms with adjustable internal loads. Tests were also conducted in the same building on a third air handler serving areas used for instruction and by building staff. One of the two fault detection and diagnosis (FDD) methods used first-principles-based models of system components. The data used by this approach were obtained from sensors typically installed for control purposes. The second method was based on semiempirical correlations of submetered electrical power with flow rates or process control signals. Faults were introduced into the air-mixing, filter-coil, and fan sections of each of the three air-handling units. In the matched air-handling units, faults were implemented over three blind test periods (summer, winter, and spring operating conditions). In each test period, the precise timing of the implementation of the fault conditions was unknown to the researchers. The faults were, however, selected from an agreed set of conditions and magnitudes, established for each season. This was necessary to ensure that at least some magnitudes of the faults could be detected by the FDD methods during the limited test period. Six faults were used for a single summer test period involving the third air-handling unit. These fault conditions were completely unknown to the researchers and the test period was truly blind. The two FDD methods were evaluated on the basis of their sensitivity, robustness, the number of sensors required, and ease of implementation. Both methods detected nearly all of the faults in the two matched air-handling units but fewer of the unknown faults in the third air-handling unit. Fault diagnosis was more difficult than detection. The first-principles-based method misdiagnosed several faults. The electrical power correlation method demonstrated greater success in diagnosis, although the limited number of faults addressed in the tests contributed to this success. The first-principles-based models require a larger number of sensors than the electrical power correlation models, although the latter method requires power meters that are not typically installed. The first-principles-based models require training data for each subsystem model to tune the respective parameters so that the model predictions more precisely represent the target system. This is obtained by an open-loop test procedure. The electrical power correlation method uses polynomial models generated from data collected from “normal” system operation, under closed-loop control. Both methods were found to require further work in three principal areas: to reduce the number of parameters to be identified; to assess the impact of less expensive or fewer sensors; and to further automate their implementation. The first-principles-based models also require further work to improve the robustness of predictions.