A fault diagnosis strategy for refrigerant leakage of the air conditioning system in high-efficiency internet data centers

Abstract

Internet data centers (IDCs) are large energy consumers and the IDCs air conditioning system will inevitably experience refrigerant leakage due to long-term and non-stop operation, which increases the risk of computer services’ health and leads to unnecessary energy waste. Therefore, this paper presents a fault diagnosis strategy for refrigerant leakage of the IDCs air conditioning systems based on deep neural network (DNN). The Gini coefficient is utilized to choose important feature variables. Then a fault diagnosis model was developed based on the DNN. On-the-spot experiments of an IDCs air conditioning system are conducted to collect practical operational data to validate this strategy. Refrigerant charge under normal conditions and five various leakage levels are investigated. The offline data of IDCs air conditioning system are collected to train the DNN models, testing results show that the proposed DNN model has good classification performance and generalization ability. The accuracy, geometric mean accuracy (GMA), false alarm rate (FAR), missing alarm rate (MAR) reach to 99.99%, 99.92%, 0%, 0%, respectively. A small amount of online data was used to update the model, the classification performance of the model will be greatly improved, which shows that the proposed DNN model has great potential for online data classification. Accuracy increase by 26.62%, from 73.66% to 93.27%, FAR decrease from 32.82% to 0%.