Research On Chiller Fault Detection And Diagnosis Based On Deep Learning

The chillers are the main energy-consuming components of heating,ventilation and air-conditioning(HVAC)systems.When the chiller fails,it will not only consume a lot of energy,but also increase the cost of operation and maintenance,reduce indoor comfort,and even damage the equipment,resulting in safety accidents.Applying the fault detection and diagnosis(FDD)method to the field of chillers,discovering chiller faults in time and determining the type of failure,so as to make decisions to eliminate the faults,has important economic and engineering significance.This study divides the FDD method of chiller into three parts: data preprocessing,discriminant analysis,evaluation and decision-making,and optimizes the first two parts based on deep learning technology.In the process of data preprocessing,with the goal of reducing feature dimensions and improving efficiency,a feature selection method based on non-parametric kernel density estimation(KDE)and a features extraction method based on factor analysis(FA)are proposed;In the discriminant analysis,aiming at improving the accuracy of the model,a bayesian network(BN)based on prior probability update(UP)mechanism and non-parametric kernel density estimation(KDE)is proposed :(1)This method uses the characteristics of BN to deal with uncertain problems,and uses BN as the discriminant analysis model of the chiller FDD method,and realizes the integration of the fault diagnosis and detection process on the BN structure;(2)The method uses the fusion ability of BN,the UP and KDE are integrated into the BN on the parameters to optimize the a priori probability and conditional probability distribution,wherein UP updates the a priori probability in real time by setting the sample update capacity,and gradually mines the a priori information of the sample,it is conducive to the combination of the actual field and the model.KDE estimates the true distribution of the sample,avoiding the uncertainty caused by the assumption of the sample distribution.This study uses ASHRAE RP-1043 experimental data to verify the proposed method.The results show that after KDE feature selection and FA feature extraction,the feature parameter dimension is reduced from 64 to 5 dimensions,which greatly reduces the amount of calculation and reduces the dimension The last five features can reflect the95.145% information of the original feature;The UP-KDE-BN model’s false alarm rate(FAR)fell to 1.8%.Among the missed detection rate(MDR)and wrong diagnosis rate(WDR)indicators of various faults,the maximum MDR is 10% of refrigerant leakage(RL)failure,and the maximum WDR is 3.8% of RL,both controlled within acceptable threshold.The correct rate(CR)of normal state and various faults is basically above 90%,the lowest is RL,but also 86.2%,so the FDD method proposed in this paper has better performance in efficiency and accuracy.