Identification of switching linear systems using self-organizing models with application to silicon prediction in hot metal

• On-line method for identification of switching linear models.
• Self-organizing learning procedure for model selection.
• Method identifies switching models accurately on simulated data.
• Application to a metallurgical process demonstrates practical feasibility.

An approach based on self-organizing models is presented for identification of piecewise linear switching systems, whose dynamics switch between a number of modes. The proposed method is based on a formulation of the identification problem as a generalized plane-clustering problem, which is solved using self-organizing maps. The method does not assume that the system modes depend on the state, but the mode switches may occur in an arbitrary and unknown manner. The procedure does not require knowledge of the number of modes or system orders, and it can be used for both on-line and off-line identification. Numerical examples illustrate that the procedure identifies switching systems correctly. The identification method is also applied to data from an industrial blast furnace for modeling and prediction of the silicon content of the hot metal. A switched linear model is demonstrated to capture different dynamics of the process and an analysis of the results reveals how mode switching models gradual and rapid changes in the output. The resulting models are finally shown to provide insight into factors that govern the silicon content in the blast furnace in different states.

Figure optionsDownload as PowerPoint slide