Research On Fault Detection And Diagnosis Methods For Chillers Based On LightGBM

Chillers are the main energy-consuming components of HVAC systems(Heating,Ventilation and Air-Conditioning,HVAC).A fault in a chiller not only wastes a lot of energy,but also increases operation and maintenance costs,reduces indoor comfort and can even cause equipment damage and safety-related accidents.The application of Fault Detection and Diagnosis(FDD)technology to chillers is of great economic and engineering importance,enabling the timely detection and removal of chiller faults and ensuring the safe and stable operation of the system.Seven common faults are identified in chillers: condenser fouling(CF),refrigerant deficiency(RL),refrigerant overload(RO),lubricant overload(EO),cooling water flow deficiency(FWC),chilled water flow deficiency(FWE)and non-condensable gas(NC).The main objective of this paper is to develop and apply an efficient fault detection and diagnosis(FDD)method to chillers.The method combines the integrated decision tree based learning algorithm LightGBM with an exponentially weighted moving average control chart(EWMA control chart)and aims to improve the accuracy of the baseline model and increase the efficiency of detection of low level faults.The process of the chiller FDD method developed in this paper is divided into two steps: benchmark model training and fault detection and diagnosis.The three parameters Evap Tons(evaporative load),TEO(chilled water outlet temperature)and TCI(cooling water inlet temperature)of the normal operation data are selected as inputs to the model,and five characteristic parameters of the normal operation data are used: TEI-TEO(chilled water inlet and outlet temperature difference),TCO-TCI(cooling water inlet and outlet temperature difference),TO＿sump(oil tank temperature),TRC＿sub(subcooling)and TCA(condensing approximation temperature)were used as model outputs,and the normal steady-state data were used to build a regression model of the chiller characteristics after data pre-processing.After obtaining the regression model,the residuals between the predicted and actual values of the characteristic parameter model are analysed and the residual range of the normal data is determined using the EWMA method,i.e.the fault thresholds(upper and lower thresholds)are determined.This is followed by fault detection and diagnosis,observing whether the characteristic parameter residuals are within the range of the fault thresholds,determining whether a fault has occurred,and then using the fault diagnosis rule table to determine the fault category.The proposed FDD method was validated using the ASHRAE RP-1043 project dataset and operational data from a chiller condenser cleaning before and after the actual application in the project.The results show that the LightGBM algorithm improves the accuracy of the baseline model and the EWMA control chart improves the efficiency of detection of low-level faults.Fault detection rates of 85.19%,75.93%,100%,91.67,100%,100% and 100% were achieved for seven types of faults: CF,RL,RO,EO,FWC,FWE and NC respectively.The fault detection rates for CF,RL and EO at the three lower fault levels still need to be improved,with 55.56%,11.11% and 66.66% detection rates for the three fault first fault levels respectively.The validation results for actual operating chillers show that the proposed FDD method for chillers can effectively detect and diagnose the fault of condenser fouling in chillers.