Fault Detection And Diagnosis In Chillers Based On KECA-ELM

Heating,ventilating and air conditioning(HVAC)systems whose structures are complicated are used increasingly in buildings.Faults will inevitably occur in long-term operation under variable working conditions.Operating faults will lead to energy consumption,lower operating efficiency and users' discomfort.Chillers are one of the most widely used HVAC systems in buildings.The application of fault detection and diagnosis(FDD)in chillers is beneficial to guarantee the best operational conditions of complex HVAC systems with the minimum cost.This study proposes a FDD strategy based on kernel entropy component analysis(KECA)and extreme learning machine(ELM).It has no Gaussian assumption and is effective for nonlinear process modeling.This method is effectively applied to seven common faults in a typical chiller from ASHRAE(American Society of Heating,Refrigerating and Air Conditioning Engineers)research project 1043(RP-1043).The specific work is presented as follows:(1)Considering the inadequacies when using Principle Component Analysis(PCA)to deal with the nonlinear data,kernel principal component analysis(KPCA)and the KECA were introduced to detect the operating faults.As compared to the KPCA,the principal components selected by the KECA with Renyi quadratic entropy can ensure the minimum information loss during dimensionality reduction.Using the seven single faults of the chiller,the KECA was validated.Results showed that the KECA-based fault detection model had a better fault detection performance compared with PCA and KPCA.(2)The traditional indices of Hotelling's T2 statistic and Squared Prediction Error(SPE)statistic are based on Gauss hypothesis,thus there are false alarms.After obtaining the principal component based on KECA,it is found that the projected data has an angular structure,and thus the fault data and normal data can be identified.Therefore,a new statistic index based on Cauchy-Schwarz(CS)is proposed to express the similarity between different probability distributions of data.This method avoided the Gaussian assumption.The statistical monitoring control limit was determined by Kernel Density Estimation(KDE).Considering the difficulties in smooth factor selection in KDE,an improved gray wolf optimization(IGWO)was used quickly and accurately to optimize smooth factor.Compared with GWO,IGWO improved the balance between local search and global search,had faster convergence and fewer chances to get stuck at local minima.The simulation results showed that this method had a better detection performance.(3)Fault detection is only to determine whether a fault has occurred,and what kind of fault should be classified by pattern recognition.In this paper,ELM was introduced for fault identification.The hidden layer selection has a certain influence on the ELM,which is usually located in a small range.Extreme learning machine based on voting rules was proposed to solve this problem,i.e.Voting based extreme learning machine(VELM).Its basic idea was to divide ELM into several sub-ELMs to improve the stability of the classifier without increasing the training time.To improve the fault diagnosis performance,this study proposed a KECA-VELM based method,which integrates the VELM method with the KECA method.The seven faults in the chiller were used to analyze the proposed fault diagnosis performance.