Impacts of data uncertainty on the performance of data-driven-based building fault diagnosis

Abstract Artificial intelligence shows powerful capacity in modern building fault detection and diagnosis (FDD) systems. Existing data-driven-based FDD models are typically developed based on experimental data or simulation data and their performance highly depend on the data quality. Nevertheless, the available data in real buildings are often uncertain due to innate measurement errors, discontinuous sampling, sensor aging, etc. Accordingly, existing data-driven-based approaches may not present sufficient values for practical applications. This paper investigates the impacts of data uncertainty on data-driven-based building FDD model from two levels. At the high level, the impacts of data uncertainty on performance of FDD model are examined under different input feature numbers, i.e., 3, 8, 22 and 36. At the low level, the impacts of data uncertainty under various input features are evaluated at four uncertainty levels. The Coefficient of Performance Degradation (CPD) index is designed to quantify the impacts of data uncertainty. The chiller operation data of the ASHRAE 1043-RP project is employed as a case study. Diagnosis results reveal that the performance of FDD model will decline as the reduction of feature number under uncertain inputs. Besides, the diagnostic accuracy dramatically descends with increases of uncertainty level. The maximum accuracy decline is 20.9% when uncertainty level of kW increased from ±0.0 to ±1.0. Moreover, results show the data uncertainties cause greater effects on diagnosing global fault rather than diagnosing local fault.