Order reduction of large scale second-order systems using Krylov subspace methods (Q2491707)

Krylov subspace methods based on moment matching are used widely for order reduction of large-scale linear time-invariant systems, but they are customarily applied only to first-order systems. In many applications -- for example, in electrical and mechanical systems -- models consist of second-order differential equations, and Krylov methods cannot be applied directly. In this paper the authors propose two methods of adapting the standard approach to make Krylov methods work for second-order systems. The object of Krylov subspace methods is to reduce the order of the linear system corresponding to the differential equation models. Solution of unreduced high order systems can be computationally impractical, and moment matching methods are among the best techniques known for order reduction. The idea is to construct a reduced, computationally tractable model whose moments match the unreduced system up to a certain order. The higher-order models considered here have the form \[ M\ddot z(t)+ D\dot z(t)+ Kz(t)=\overline b u(t), \] \[ y(t)= \overline c^T z(t), \] where \(M\), \(D\) and \(K\in \mathbb{R}^{n\times n}\), \(\overline b\) and \(\overline c\) are constant column vectors, \(u\) is the input, \(y\) is the output, and \(z(t)\) is the state vector of the system. This system is recast in consolidated form for convenience as \[ E\dot x(t)= Ax(t)+ bu(t) \] \[ y(t)= c^T x(t), \] where \[ E= \left[\begin{matrix} I & 0\\ 0 & M\end{matrix}\right],\quad A= \left[\begin{matrix} 0 & T\\ -K & -D\end{matrix}\right],\quad x(t)= \left[\begin{matrix} z(t)\\ \dot z(t)\end{matrix}\right], \] \[ b= \left[\begin{matrix} 0\\ \overline b\end{matrix}\right]\quad\text{and}\quad c= [\overline c^T\;0], \] and the so-called moments are defined by \[ m_i= c^T(A^{-1} E)^i A^{-1}b,\quad i= 0,1,2,\dots\;. \] The scalar values \(m_i\) are the negative coefficients of the Taylor series expansion around zero of the system's transfer function. Both methods proposed by the authors proceed by moment matching in state space. The first reduction method identifies projection matrices to apply to the second-order models. This method preserves the second-order structure of the original model. The second reduction method applies projections instead to the state space model. It also preserves the second-order structure of the model but it achieves a higher level of moment matching. The authors provide a numerical example based on a second-order clamped beam model of dimension 174 with one input and one output. They apply and compare the performance of their two reduction methods.