With increased emphasis on plant automation and the implementation of advanced process control systems in the mineral processing industries, the need for accurate process models has become greater than ever. These models are often developed or identified from historic process data, as the process systems may be too complex to model from first principles. Under these circumstances, identification can be a challenging task and in the process industries the problem is complicated considerably by the presence of noise from various sources, nonstationarity of the data and intermittence, such as observed in particulate flows. To complicate matters, it may not be possible to build satisfactory models with the aid of traditional methods, such as frequency analysis and linear modelling either. Similar problems arise with the application of nonlinear theory developed over the last few decades, where much of the analysis depends on embedding of the data in a phase space or pseudophase space, since these methods were not originally designed to deal with noisy systems. In contrast, most of these disadvantages can be surmounted by use of singular spectrum analysis. It allows the time series to be decomposed into different components, e.g. the underlying signal itself, as well as various noise components, which can subsequently be removed from the data. As is shown in this paper, removal of the minor components of the data can lead to significant improvement in the identification of the system. Four strategies are considered, viz. system identification with models fitted to the original data, models fitted to the smoothed data, an assembly of models fitted to the components of the time series and smoothing of the data based on multivariate embedding of time series. (C) 2003 Elsevier Ltd. All rights reserved.