Nonlinear state estimation using variable-structure neural network

State estimation is an integral part of process monitoring, diagnosis and c ontrol. Due to the mathematical complexity of nonlinear model, optimal stat e estimation is much less established in practice. In this work, the variab le-structure neural network (VSNN) of Tung et al. is employed for nonlinear state estimation. Similar to the Kalman filter approach, the filter gain i s adjusted according to the the ratio of noise and error covariance. An alg orithm is proposed to approximate the error covariance and, subsequently, r esults in a non-iterative procedure in the filter gain calculation. Moreove r, when some of the states are not measurable, the VSNN naturally results i n a RecN-like architecture for the unmeasured states. A chemical reactor ex ample is used to demonstrate the effectiveness of the VSNN state estimation scheme. Results show that the variable-structure neural network provides e ffective state estimation for nonlinear chemical reaction systems. More imp ortantly, ail the on-line computing can be carried out with standard neural network software in a non-iterative manner.