（天津大学 机械工程学院， 天津 300072）
Novelty detection based on Gaussian mixture models in ICA space
PEI Zhijun, TAO Jianhua
(School of Mechanical Engineering, Tianjin University，Tianjin 300072, China)
Abstract: A novelty detector learns the model of normality in the training stage using only normal samples and abnormalities are then identified by testing for novelty against that model. Gaussian mixture models can be used to model data general distributions for novelty detection. But given high data dimensionality, a very large number of training samples are needed for modeling, there are also too many free parameters. ICA is a subspace projection technique that can project data from a highdimensional space to a lowerdimensional space by computing independent components of the data. So this paper proposed a novelty detection based on Gaussian mixture models in ICA space, which could improve the dimension curse problem and decrease the free parameters. The method is verified by the experiments. ......