One-Class Support Vector Machine for Functional Data Novelty Detection

Abstract

Novelty detection builds a model only with a large number of normal samples to detect unknown abnormalities. Based on the kernel theory and the optimization method, One-Class Support Vector Machine (OCSVM) can build a high-performance detection model with only a small part of training samples. As a result, OCSVM has become a very popular novelty detection method. However, with the increasing of the sensor precision and the data acquisition frequency in large-scale complex production processes, the collected data present high-dimension and more complex trend. Each data shows obviously functional nature (called functional data). Therefore, How to deal with these functional data and to dig out the production performance messages in them brings a new challenge to novelty detection. For this purpose, this paper proposes an OCSVM algorithm based on Functional Data Analysis (FDA), which is called Functional OCSVM. The experimental results show that Functional OCSVM can achieve better detecting results than original OCSVM by using the functional nature of data.