Fault detection and diagnosis using a probabilistic modeling approach.

摘要

The building sector in the U.S. accounts for approximately 40% of the national primary energy usage, 37% of which comes from space heating. Faulty systems and control schemes can increase energy usage by 15% to 20%. Quick and accurate fault detection and diagnosis will play a key role in reducing building energy consumption.;This thesis explores potential Bayesian fault detection, diagnosis and correction methods in commercial buildings. The first experiment investigated fault correction in a test building. A test-building model was calibrated to measured data. It was then found, that by implementing a nighttime setback, energy savings of approximately 20 percent could be achieved.;Next, experiments were carried out using surrogate model data to investigate a number of hydraulic system faults, such as an inefficient boiler, high valve leakage, valves with high hysteresis and heat exchanger fouling. It was determined, that experimentally with surrogate data, Bayesian methods are effective for detecting hydraulic heating system faults.;Bayesian methods were then used to examine heat exchanger fouling for two heat exchangers in a test building using measured data. The amount of propagated model uncertainty and measurement noise present made fault detection in this case more difficult. The heat transfer values in the heat exchangers were not determined to be low enough to be considered significantly faulty.;Lastly, an experiment was carried out to test if the number of models created could be minimized. The models were correct in predicting faults in many circumstances, however mischaracterized faults were not uncommon.