Pattern recognition-based chillers fault detection method using Support Vector Data Description (SVDD)

Abstract

In building management systems, the fault-free data are usually available while the fault data are generally insufficient. The fault detection task can be considered as a typical one-class classification problem, which aims to differentiate the fault-free data class from all other possible fault data classes. In this study, a pattern recognition-based chiller fault detection method is proposed using a novel one-class classification algorithm, i.e. Support Vector Data Description (SVDD). Its basic idea is to find a minimum-volume hypersphere in a high dimensional feature space to enclose most of the fault-free data. When a fault occurs, the fault data will be outliers of the hypersphere. Compared with Principle Component Analysis (PCA), it has no Gaussian assumption and is effective for nonlinear process modeling. It can also compress process variables of wide-range operating conditions into a single model with higher fault detection accuracy. This method is validated using the experimental data from ASHRAE Research Project 1043 (RP-1043). Results show that the SVDD-based method has better chiller fault detection performance compared with PCA-based methods.