A fast SVDD algorithm based on decomposition and combination for fault detection

Abstract

SVDD is an effective tool for novelty detection. But due to space complexity of matrix operations, the optimization process using original support vector data description (SVDD) algorithm becomes memory and time consuming when the size of training set increases. We present a fast SVDD algorithm based on the strategy of decomposition and combination. First, we reduce the space complexity by breaking the training dataset into subsets at random and apply SVDD to each subset. Then, based on two lemmas of random sampling and SVDD combining, we merge the data descriptions into common decision boundary. We repeat the above two-step until achieving description of the entire data sample. Experimental results show that the algorithm is more superiority than original SVDD algorithm in achieving the sample description, especially on the large scale sample dataset.