Fault Diagnosis In Chillers Based On XGBoost

At present,the energy consumption of buildings accounts for up to 30% of the total energy consumption of the society.While the energy consumption of heating,ventilation,air conditioning and refrigeration(HVAC&R)systems accounts for more than half of the building energy consumption.As the core equipment of the HVAC&R system,once chillers fail will not only cause low system operation efficiency,increase energy loss,but also lead to loss of system functions,which will endanger life and property.So,it is of great significance to carry out fault diagnosis research on chiller units to discover unit failures in time,and it has become one of the hottest research directions in the field of energy sources.This paper combined with ASHRAE RP-1043 chillers failure simulation experimental data,using analytical methods to study the chillers operation monitoring data.Founding some questions like imbalance,noise,non-linearity,non-Gaussian,and high feature dimension in chillers operation monitoring data.Aiming at these problems,the following research is based on extreme gradient boosting(XGBoost)algorithm which is conducted on the chillers fault diagnosis:(1)The XGBoost algorithm is used for chillers fault diagnosis task,and a three-level evaluation system based on the confusion matrix is established.And there are 22 datasets were established based on the RP-1043 fault simulation database.Compared with support vector machine and artificial neural network algorithms,the XGBoost algorithm average Kappa is 0.912 when the number of samples is small.When the sample size is large,the average Kappa is 0.931,which is higher than other algorithms,which proves that the XGBoost algorithm is an effective solution for chiller fault diagnosis.(2)A feature selection strategy is proposed based on the principle of decision tree partitioning when training the model with the XGBoost algorithm.From the 64-dimensional features of the Data8 dataset(Level 1),23-dimensional features which are more important to the XGBoost diagnostic model are selected.Compared with other feature selection schemes,the method proposed in this paper has the highest accuracy rate in the same dimension,and improves the training speed of the chiller fault diagnosis model without affecting the model diagnosis performance.(3)Improving the particle swarm optimization(PSO),The more swarms particle swarm optimization(MSPSO)algorithm was proposed to optimize the XGBoost hyper parameters,and a MSPSO-XGBoost composite model was established and applied to the diagnosis of chillers incipient faults(Level 1 fault).The experimental results show that the MSPSO-XGBoost composite model can improve the accuracy rate of chiller fault diagnosis to 99.67%,reduce the false negative rate and false alarm rate of the system,and is suitable for the chiller incipient fault diagnosis task.(4)For data imbalances,the minority oversampling under local area density(MOLAD)algorithm was proposed to establish a MOLAD-XGBoost imbalance data fault diagnosis model which is based on synthetic minority oversampling technology(SMOTE).Comparing with SMOTEXGBoost,the improved MOLAD-XGBoost is good at synthesizing new samples with higher quality for data imbalance,increases the average F1 from 0.9481 to 0.9511,which is more suitable for the imbalance data of chillers fault diagnosis;Based on cost-sensitive learning,sensitive weights are proposed and verified.The results show that setting appropriate sensitive weights can improve the recall rate for specified faults at the cost of slightly reducing the accuracy of the classifier,which has certain practical value.