Refrigerant Leakage Detection For IDC Air Conditioning Systems Based On Associative Classification

In recent years,with the development of big data and cloud computing technologies,the energy consumption of Internet data centers(IDC)is getting higher and higher.The energy consumption of air conditioning systems used to maintain the stability of IT equipment accounts for up to 40% of total energy consumption in IDC.One of the most common failures in air conditioning systems in IDC is the fault of refrigerant leakage.The fault of refrigerant leakage,on the one hand,leads the air conditioning system to deviate from the optimal operating area,reduce its operating efficiency and causes waste of energy consumption;on the other hand,reduces the cooling capacity,affects the temperature and humidity in IDC,and thus harms the safety of the IT equipment.Therefore,the detection of refrigerant leakage of IDC air conditioning system is of great practical significance.Aiming at the refrigerant leakage fault of IDC air conditioning system,this thesis proposes a fault detection strategy based on association classification.The main ideas of constructing a fault detection model are as follows: Firstly,a training item set is obtained through preprocessing the training data,and when the minimum support and the minimum confidence threshold is set,association rules can be mined by using the FP-growth algorithm.Secondly,the class label rules whose right-hand side is the refrigerant leakage fault are extracted out of the association rules.Then,according to the rule rank principle the class label rules are sorted in descending order,and redundant rules are pruned so as to maintain the pruned class label rules.Finally,the training data set is scanned multiple times and compared to the pruned class label rules according to the association classification,and only limited number of rules are selected for fault detection,named classifier rules.At the meanwhile,a default fault class is deduced.Consequently,the fault detection model upon the association classification is developed,which consists of classifier rules and a default fault class.It can be then used for the fault detection of refrigerant leakage in IDC air conditioning systems.In this thesis,experiments on an IDC air conditioning system was conducted to collect operational data under five different refrigerant charge levels,including normal refrigerant and 10%,20%,30%,and 40% leakage faults,respectively.A fault detection model is developed and validated to test its performance.The experimental results show that the fault detection accuracy of the model for the five different refrigerant charge levels is 95.4%,93.5%,88.2%,90.9% and 99.6%,respectively,with the average accuracy up to 93.52%;wrongly detected faults are recognized as the fault closest to the real fault.The fault detection model based on the association classification has a very satisfying performance on the fault detection for refrigerant leakage in IDC air conditioning systems.The fault detection model based on association rules may have fault bias problems due to data imbalance.This thesis proposes an improved fault detection strategy by employing data resampling methods.Four kinds of resampling methods,such as SMOTE,SMOTESVM,ENN and SMOTEENN,are used to optimize the three imbalanced data sets.It is found that the four data resampling methods are all beneficial for improving the fault detection model,and the data resampling methods of SMOTE,SMOTESVM and SMOTEENN have better optimization effects than ENN.