A New Detection And Diagnosis Method Research For Chillers Based On Nonlinear Support Vector Regression

The heating, ventilation and air conditioning (HVAC) systems have played anincreasingly important role in people’s life and work. However, as a major componentof building energy consumption, it also greatly increases the energy consumption ofthe building. According to statistics, the energy consumption of chiller accounted for40%to50%of the total energy consumption of air conditioning systems.When thechiller fails, not only affect the safe operation of the unit and the thermal comfort ofthe indoor space, but also increases the energy consumption of the entire system,resulting in energy waste. Thus, real-time monitoring chiller running, and usingreliable fault detection and diagnosis (FDD) strategy to prevent the occurrence offailure and ensure the efficient operation of the chiller has always great practicalsignificance.This paper presents a new fault detection and diagnosis method for centrifugalchillers of building air-conditioning systems, which bases on nonlinear support vectorregression (n-SVR), and mainly consists of offline model training and online chillerfault detection and diagnosis. In the new strategy, the nonlinear support vectorregression is adopted to develop the PIs reference models, which improved theaccuracy of multiple linear regression model, when it is adopted to study a typicalnon-linearly system. The exponentially weighted moving average (EWMA) controlcharts are introduced to detect faults in a t-statistic-based way to reduce the TypeⅡerrors and improve the low-level fault detection rate. The evaporator cooling load, theleaving chilled water temperature, and the entering cooling water temperature areselected as the input variables of the reference model. Five chiller characteristicparameters that have strong sense of the physical are selected as the fault performanceindexes. After data processing, the steady-state and fault-free data are used to create acentrifugal PIs reference model. In this study, seven classical faults of the centrifugalchillers are chosen to exam FDD methods. To simplify fault rules of the traditionalFDD methods, according to the most sensitive parameter method，we can calculatedthe sensitive parameters that are most intense sensitivity to changes in the mostcharacteristic parameters when a fault occurs, and make it as the fault onlinediagnostic classifier.The new FDD strategy is respectively validated using the27sets of experimental data from ASHRAE RP-1043project and actual operating data of condenser cleanedfrom a real building HVAC system in Hong Kong. Test results show that, compared toa traditional method using multiple linear regression (MLR) and t-statistic, the newFDD method that combined use of n-SVR and EWMA can achieve significantimprovement in accuracy and reliability.