NEURAL-NETWORK-BASED SELECTION OF DYNAMIC SYSTEM PARAMETERS

The present paper explores the use of a global response technique utilizing a neurocomputing paradigm to the solution of an inverse eigenvalue problem. For a given vibratory system, the problem is one of determining a set of physical construction parameters such as mass, stiffness, and damping, to yield a desired set of eigenvalues and eigenvectors, In absence of an exact analytical solution to the problem, approximations yielded by the backpropagation and an improved counterpropagation neural networks are examined. While estimates from both network architectures are acceptable for technical implementation, the CPN network is much simpler to train, Improvements to the CPN network include a dynamic adjustment of the network size, the use of averaging operators for training, and an increased accuracy of approximations based on a nonlinear blend of interconnection weights. A state matrix representation of a truck suspension system is used as an illustrative problem to demonstrate the effectiveness of the neurocomputing approach in such inverse eigenvalue problems.