Font Size: 

Many real-world problems involve segmenting temporal data into homogeneous regimes in order to extract relevant features. This operation consists in automatically grouping the points of a single series into clusters associated with contiguous or locally contiguous time intervals. In this work, we are mainly interested in discovering segments that reflect changes in a signal derivative, which are generally associated with dynamic phenomena governed by physical laws. This involves revisiting classical approaches based on mixture models or hidden Markov models. The segmentation approach proposed in this paper is thus inspired by the family of structural time series models. It is an extension of the local level model where the first derivative of the trend component is no longer distributed according to a simple Gaussian distribution, but can switch between different Gaussians via a hidden Markov chain. The resulting model structure, with two levels of latent variables (the denoised unknown trend of the series and its hidden states indicating segment membership), belongs to the family of state-space models. A variational Expectation Maximization (VEM) algorithm is proposed for maximum likelihood estimation. The method is implemented on simulated series and also on real-world series from an energy efficiency context.

model-based segmentation, time series, structural model, local level model with drift