Change point detection of time series based on relevance vector machine and Bayesian framework with application to steel manufacturing

Abstract. The change point detection of time series is an urgent issue in the continuous casting quality control. A novel method based on Relevance vector machine (RVM) in the Bayesian framework is proposed for change points detection. First, the posterior distribution of run length is introduced into the change point detection framework. Second, RVM is improved to calculate the predicted distribution of the observation data, which is embedded in the detection framework to achieve the posterior distribution. The posterior probability of the maximum run length is calculated to describe the severity of the data change. Then, the reprocessing is proposed to modify redundant change points in local time. Eventually, traditional Bayesian and Singular Spectrum Transforms are used for comparison, and the effectiveness and superiority of the RVM-Bayesian are illustrated by the continuous casting process. The results show that RVM-Bayesian method can not only accurately detect the change points in the time series, but also characterize the severity of the change points.