Classification of multiple time signals using localized frequency characteristics applied to industrial process monitoring

Abstract

A general framework for regression modeling using localized frequency characteristics of explanatory variables is proposed. This novel framework can be used in any application where the aim is to model an evolving process sequentially based on multiple time series data. Furthermore, this framework allows time series to be transformed and combined to simultaneously boost important characteristics and reduce noise. A wavelet transform is used to isolate key frequency structure and perform data reduction. The method is highly adaptive, since wavelets are effective at extracting localized information from noisy data. This adaptivity allows rapid identification of changes in the evolving process. Finally, a regression model uses functions of the wavelet coefficients to classify the evolving process into one of a set of states which can then be used for automatic monitoring of the system. As motivation and illustration, industrial process monitoring using electrical tomography measurements is considered. This technique provides useful data without intruding into the industrial process. Statistics derived from the wavelet transform of the tomographic data can be enormously helpful in monitoring and controlling the process. The predictive power of the proposed approach is explored using real and simulated tomographic data. In both cases, the resulting models successfully classify different flow regimes and hence provide the basis for reliable online monitoring and control of

Spatio-temporal forecasting using wavelet transform-based decision trees with application to air quality and covid-19 forecasting

“…Thus, when j is small, wavelets are highly localized at a fine scale resolution level, representing brief transient effects. Conversely, when j is large, wavelets represent lower frequency activity at coarser scale resolution levels (Aykroyd et al, 2016). Here the factor of 2 j/2 ensures energy preservation, defined by…”

Dynamic detection of anomalous regions within distributed acoustic sensing data streams using locally stationary wavelet time series

“…scope of this article, but popular classification methods include hidden Markov models (HMM) (see e.g. Rabiner (1989); Ephraim and Merhav (2002); Cappé et al (2009)); support vector machines (Cortes and Vapnik 1995;Muller et al 2001;Kampouraki et al 2009); Gaussian mixture models (McLachlan and Peel 2004;Povinelli et al 2004;Kersten 2014); nearest neigbour classifiers (Zhang et al 2004;Wei and Keogh 2006) and multiscale methods (Chan and Fu 1999;Mörchen 2003;Aykroyd et al 2016) to name but a few. More recent contributions for large-scale (online) classification include the MOA machine learning framework (Bifet et al 2010;Read et al 2012).…”