Predictive temporal patterns detection in multivariate dynamic data system

In this paper we present a method for detecting multivariate temporal patterns that are characteristic and predictive of significant events in a multivariate dynamic data system. A new hybrid RPS-GMM method is applied to identify patterns. This method constructs phase space embedding by using individual embedding of each variable sequences. We employ discriminative approach by applying Gaussian Mixture Model (GMM) to the multivariate sequence data to cluster multi-dimensional data into three categories of signals, e.g. normal, patterns and events. An optimization method is applied to the objective function to search an optimal classifier to identify temporal patterns that are predictive of future events. We performed two experimental applications using chaotic time series and Sludge Volume Index (SVI) series related to the Sludge Bulking problem. Experiments show that the new approach presented here significantly outperforms the original RPS framework and neural network method.