A cumulant based algorithm for the identification of input-output quadratic systems (Q5960311)

The authors consider the identification of the following quadratic system: \[ \begin{multlined} z(n)=\sum_{i=1}^{k_1} a_i z(n-i)+\sum_{i=0}^{k_2} b_i u(n-i) +\sum_{i=1}^{k_3}\sum_{j=1}^{k_3} A_{ij} z(n-i)u(n-j)\cr +\sum_{i=0}^{k_4} \sum_{j=1}^{k_4}B_{ij} u(n-i)u(n-j)+\sum_{i=1}^{k_5}\sum_{j=1}^{k_5} c_{ij}z(n-i)z(n-j),\end{multlined} \] \( y(n)= z(n)+\eta (n),\) where \(u(n)\) and \(y(n)\) are the input and the measured output corrupted by the additive measurement noise \(\eta (n)\), which is a zero mean and independent of \(u(n)\). The input excitation \(u(n)\) is a zero mean white stationary Gaussian stochastic process. They propose an iteration algorithm to estimate the coefficients \(a_i\), \(b_i\), \(A_{ij}, B_{ij}\), and \( C_{ij}\) using input-output cross cumulants up to the fifth order, thus, the identification problem is reduced to the solution of successive triangular linear systems of equations. Simulation studies are presented, and the proposed method is compared with two least squares type identification algorithms. Their cumulant based algorithm is tested with real data produced by a robotic manipulator.