Information fusion method for system identification based on sensitivity analysis

We considers the identification of a parametrized time-invariant nonlinear plant using a smooth model such as a sigmoid nonlinear network. There is measurement noise associated with the plant parameters as well as with its input and output. An initial plant model is obtained by utilizing the domain-specific knowledge in terms of the fundamental plant equations, which in general only partially capture the plant dynamics. Once the initial model is fixed, measurements are collected on the plant parameters and input/output. We show that the independently and identically distributed (iid) measurements can be fused with the initial plant model by recomputing the parameters. The updated parameters yield a more accurate identifier of the original plant both in the parameter space and the input/output space. The method is based on empirical versions of the closed-form solutions derived in the nuclear engineering literature for an ideal version of the problem based on sensitivity analysis. We show the asymptotic convergence of our computational procedure as well as derive its finite sample results. We illustrate the method using an identifier based on a sigmoid feedforward neural network.