Effects of sensor measurement error on fault detection and diagnosis model for data center composite cooling system

Abstract

Fault detection and diagnosis (FDD) model for the cooling system is beneficial in elevating the reliability of data centers. Nevertheless, the model accuracy could be degraded by sensor measurement error, which may arise due to environmental interferences or inadequate maintenance practices. In the study, the impacts of sensor measurement error on the convolutional neuron network (CNN) based FDD model for the data center composite cooling system are assessed. Additionally, the coupled effects of sensor error and system control strategies on the FDD model are investigated. The results indicate that in vapor compression mode, a negative fixed sensor error of 1 K leads to an average 5 % greater decline in the CNN model accuracy compared to a positive error of the same magnitude. In contrast, the positive fixed error causes a 6.5 % higher decrease in heat pipe mode. Additionally, sensor errors have a negligible impact on model accuracy until exceeding the threshold, and the threshold of fixed error is 0.2 K in CNN model. Further, as a key control strategy involved parameters, the evaporating temperature error is critical to FDD model accuracy. In the fixed bias conditions, when the error magnitude is 1 K, the accuracy of FDD model decreases within the range of 24.8 % to 45.1 %.