Novelty detection examines problems where knowledge about abnormal conditions is not available. It is especially useful for monitoring safety-critical systems in which abnormal conditions rarely occur and knowledge about abnormality in such a system is usually very limited or null. It builds a detector using available data collected and knowledge captured from all known conditions. The detector is then used to surveil the target system. If the output of the detector is significantly different from that of known conditions, then it warns that a novel event has occurred in the system.
A reliable novelty detector should have a closed decision surface. Therefore, the project will focus on techniques that construct closed decision surface, and minimize the volume of the surface. It will develop a method that extracts critical information by identifying extreme patterns surrounding the given pattern set. Factors that control detector performance will be optimised. The performance of the developed novelty detector will be evaluated on benchmark datasets along with the comparison on the state-of-the-art novelty detection approaches.