Study on the support vector data description (SVDD)-based chiller sensor fault detection efficiencies.

Abstract

In refrigeration and air conditioning system, sensors play important roles in recording the performance data and monitoring the operative modes. Sensor faults especially the sensor outputs biases, may have severe effects on the energy consumption and operation cost of the system caused by incorrect estimation of the operating state, inappropriate action of controlling component. These errors or biases should be detected in time to avoid further damage. According to the pattern recognition theory, the task of fault detection can be considered as a one-class classification problem. A one-class classifier, Support Vector Data Description (SVDD) algorithm has advantages on describing the nonlinear data that violate the Gaussian distribution. And it was employed for detecting sensor faults in the screw chiller system in this study. Chiller practical operating data was used to validate the method. The fault detection efficiencies were analyzed with different artificially introduced levels of sensor biases. The grid search and the 10-fold cross validation method was adopted to search for an optimal pair of (C, g) parameters in the SVDD model so as to obtain good generalization performance and avoid overfitting problems. Results show that the SVDD-based method perform well in detecting chiller sensor faults, even for low temperature sensor faults with absolute magnitude around 1°C. And this can be conducive to early fault detection and reduce the loss.