Multi-Scale Kernel Latent Variable models for nonlinear time series pattern matching

In this paper we propose a method for nonlinear time series pattern matching: "Multi-Scale Kernel Latent Variable (MSKLV) models". The pattern matching methodology includes multi-scale analysis using wavelet decomposition of time series and finding latent vectors in the kernel feature space at different scales of wavelet decomposition. Latent vectors so obtained are matched for similarity with the corresponding latent vectors obtained for time series in the historical database. The proposed methodology is applied on time series generated in the evolving stages of disturbances of Tennesse Eastman challenge problem and MSKLV models are found to be superior to Multi-scale Latent Variable (MSLV) models.