Subspace-based identification of MIMO bilinear systems

A subspace-based algorithm for the identification of discrete-time bilinear systems is presented. The algorithm uses a subspace-based step to estimate the order of the system and to generate initial estimates of the system matrices. These matrices are iteratively refined by numerically solving a nonlinear optimization problem. The principle of separable least squares is exploited, to make the algorithm computationally efficient. The method can deal with both process and measurement noise. By means of a Monte-Carlo simulation, the method is compared with another recently proposed subspace-based algorithm for bilinear systems. It turned out that the method presented in this paper is computationally more efficient and gives more accurate models. However, sometimes the nonlinear optimization involved in the method does not converge to the global optimum and consequently the estimated model is bad.