Pulp digester level prediction using multiresolution networks of locally active units

Complex processes such as those found in some manufacturing plants have many dynamically related variables which interact to produce a system response. The goal of system identification is to construct a model of a complex process, based on observations, in order to predict future system responses and help improve its control. In this research, neural networks constructed with multiresolution locally active units are used to model the content level of a paper plant pulp digester. This problem represents a real world multidimensional dynamic system for which a predictive model does not yet exist. Two local networks are considered which use radial basis functions (RBF) and utilize the concept of multiresolution analysis (MRA). The networks differ in the way each generates an MRA candidate node set from which the most important nodes are selected for the approximation using a fast orthogonal search (FOS) algorithm. The MRA networks are compared to each other and are also shown to produce approximations using fewer nodes than those obtained using a standard RBF architecture. This report demonstrates that locally active MRA networks produce promising results for the system identification task of digester level prediction