Study On The Fault Diagnosis Method Of Chiller Based On Random Forest

Chiller is an important component and main energy-consuming equipment of Heating Ventilation and Air Conditioning(hereinafter referred to as HVAC)system with excellent reliability,stability and environmental adaptability.Faults of chiller can affect the economyefficiency,energy-efficiency and safety of HVAC system directly.In the process of industrial production,faults of chiller may present as overpressure,over temperature,liquid hammer,liquid explosion or refrigerant leakage.What's worse,it may lead to explosion,collapse,fall,fire,poisoning,suffocation,frostbite,burn,electric shock or other accidents.Therefore,the fault diagnosis technology of chiller can predict the risk of fault effectively,prevent the occurrence of catastrophic fault,improve the reliability of the system,and reduce the downtime,maintenance time and life cycle costs,which is of great significance for the safety and efficiency of chiller.Taking centrifugal chiller as a research object,this paper expounds the structure,operation mechanism and common typical faults of chiller,puts forward the fault diagnosis method and a staggered model of chiller based on Random Forest(hereinafter referred to as RF),retains the physical meaning of input variables effectively,improves the model diagnosis performance,and develops the "typical fault diagnosis platform of chiller" which provides a software platform for the engineering application of centrifugal chiller fault diagnosis.Main research job of this paper including the following four aspects:(1)It analyzes the thermodynamic characteristics of typical faults of water chillers,describes the parameter characterization,and discusses the causes and influence factors of faults.(2)Proposes a recursive feature elimination method based on RF,which can effectively select fault features,avoid the limitations of feature selection based on experience and the curse of dimensionality to solve the problem of the losing physical meaning of parameters when using conventional dimensionality reduction method due to the huge running data of chiller.Meanwhile,comparing with Support Vector Machine(hereinafter referred to as SVM)and Decision Tree(hereinafter referred to as DT),the diagnostic accuracy of RF algorithm is improved by 2.93% and 4.11%respectively.(3)Three typical fault principles with similar fault features are discussed by using confusion matrix,and a staggered model based on RF is proposed for improving accuracy by feature selection,model super parameter optimization,and data sample size expansion which improved the accuracy of the model to 98%.(4)The fault diagnosis platform of chiller is developed as "data selection","feature selection","feature analysis" and "model diagnosis" four modules in order to meet the engineering requirements,which encapsulates the data collection,storage,fault diagnosis and other processes.It providing an effective data analysis and fault diagnosis platform for the actual project.