Study On The PCA-based Sensor Fault Detection Efficiency Of The Water-cooled Chiller

Sensor fault would influence the operational situation of the HVAC system, and would strongly waste the energy consumption of the HVAC system. The sensor fault detection, diagnosis and reconstruction (FDDR) research area is a cross research area among the HVAC research area and the automation control research area. It has been concerned for recent years. Chillers are the most important energy supplying devices in the HAVC system and are the key devices that consume almost40%and more energy of the whole HVAC system. Studies on the sensor FDDR of the chiller are very important and have the theoretical and applicable values.Principal component analysis is a widely-used data analysis method for sensor FDDR. The PCA-based sensor FDDR method that uses the Q-statistic as the fault criteria is concluded in this paper. Combined the energy balance analysis with the control logic of the chiller operation, a coupled PCA model used for sensor FDDR is presented. There are8sensors included in the PCA model. They are chilled-water supply temperature (Tchws) sensor, chilled-water return temperature (Tchwr) sensor, chilled-water flow rate (Mchw) sensor, condenser-water supply temperature (Tcws) sensor, condenser-water return temperature (Tcwr) sensor, condenser-water flow rate (Mcw) sensor, the chiller electrical-power input (W) and the feedback signal (Mref) of the refrigerant flow rate control instrument, respectively. The normal PCA sensor FDDR method is validated by the field data. The fault detection efficiencies of different sensors on the different introduced fault levels are compared and validated. The results show that the fault detection efficiencies of the different sensors on the different introduced fault levels are very different. Especially on the little introduced fault level, the fault detection efficiencies are very bad. The whole fault detection efficiencies of some sensors are very low at all the introduced fault levels.Three aspects methods are presented to promote and enhanced the PCA-based sensor fault detection efficiency. They are the training matrix optimization, the measured data optimization and the detection criteria optimization.Two methods based on the concept of the space distance are presented for optimizing the training matrix. One method, the PCA-based FDD method with the Distance-based Outlier Detection (PCA-DOD), employs the Euclidean distance of the normalized original training data vectors as the detection criterion to remove the outlier form the training matrix. If the Euclidean distance’s z-score of one vector is greater than2, that sample is the outlier and should be removed from the training matrix. The other method, a self-adaptive PCA-based sensor FDD method (APCA), employs the threshold Qa of the Q-statistic as the detection criterion to remove the outlier form the training matrix. It can remove the error samples in the training matrix with a self-adaptive loop in the process of developing and training PCA model. It is the aim of the above two methods to remove the data far away from the center of the training matrix and to reduce the outlier’s influence to the PCA’s two orthogonal projection subspaces.In order to optimize the measured data, Wavelet de-noising method is employed to get rid of the noise from the training data and the new sample. Because the Wavelet transform can be decomposed by different layers, the fault detection efficiencies under different decomposed layer are presented. The result shows that the more decomposed layer employed, the more the Wavelet de-nosing PCA-based method can promote the detection efficiency higher than the normal PCA-based method.Concerned to optimize the fault detection criterion, the cross detection of three statistics, Q statistics、T2statistics and Hawkins TH2statistics, is employed. The T2statistics is the statistics of the principal component subspace. The Q statistics and the Hawkins TH2statistics are the statistics of the residual subspace. By cross detection of three statistics, the whole fault detection efficiency are promoted. The CUSUM chart is employed as an online detection method. The cumulative sum of the errors between the Q-statistics and the mean of training data Q-statistics on time is used as the criterion to detect the fault.The presented methods can enhance the PCA-based chiller sensor fault detection efficiency, and to promote the sensitivity of the sensor fault diagnosis and data reconstruction consequently.