A modified nonlinear Bayesian support vector machine (BSVM).
Technologies are needed for effectively managing building energy performance, especially in the commercial buildings space. Faults and equipment degradation can significantly increase energy consumption. Tools for fault detection and diagnostics (FDD) are needed. Most commercial buildings lack sophisticated FDD tools. The FDD tools that are gaining traction in practice are rule-based and hence, the set of faults detected and diagnosed by these tools are limited to those that are predefined. Machine learning based approaches are purely based on extracting features from historical data. They can include, but it is not necessary, assumptions from physics-based building models and expert knowledge. Therefore, a machine learning based approach will enable us to discover new conditions that have not been seen previously, which may indicate the existence of faults or equipment degradation. The BSVM has been modified to use a single class of data —a “one-class” BSVM. Traditionally, SVMs work by learning a separating line (hyperplane) between two classes of data that is maximally separated from both classes. When new data is tested, whichever side of the line it falls on determines the predicted class. The one-class BSVM instead learns the (nonlinear) boundary around the single class of training data. At test time, new data is determined to be an inlier or an outlier based on which side of the boundary it falls.