Neurocomputing based approximate models in structural analysis and optimal design.

摘要

The work entitled "Neurocomputing Based Approximate Models in Structural Analysis and Optimal Design" further explored the applicability of a neurocomputing paradigm within analysis and design of mechanical and structural systems.;It was shown that the modeling capabilities of neural networks provided a suitable technique to solve difficult identification problems such as damage detection in structural systems, and parameter selection in dynamic systems, both formulated as inverse problems. For the latter an algebraic method based on the first-order sensitivity coefficients and generalized inverse was also developed, and shown to be a viable technique for systems modeled directly in the state matrix form.;It was shown that zero-order optimization techniques such as simulated annealing are "more" tolerant of approximations in objective function evaluations, and become more efficient when estimates yielded by neural networks are used. To this extent, an improved version of the counterpropagation neural network was developed, and shown to be an adequate tool for such approximations. In addition, a new network architecture, called the z-Network, was introduced. The z-Network was shown to further improve on the accuracy of estimates due to a distance based interpolation scheme, and reduce the network operational time, due to the hierarchical structure of the network. Numerical examples show the network is a viable tool for fast function approximations (FFA).;A new technique to decompose optimization problems, which was based on the weights analysis of a backpropagation neural network, was developed. This laid the foundation for the development of a decomposition scheme to optimization of large-scale systems. In addition to partitioning of a design space, the neurocomputing strategies were used to develop selective mappings as required by the problem formulation. The z-Network shown to be an adequate tool for that purpose. In a nonlinear programming search, a penalty function strategy was used, and shown to force a rapid convergence of optimal designs in subsystems to a common point. It was also shown that the weights analysis procedure helps to design an appropriate architecture of a BPN network, and to determine the adequacy of the size of a training set.