Research And Application On Metamodel Based On Support Vector Regression For Engineering Optimization Problems

Manufacturing is a measure of a country's pillar industry, and influence national comprehensive strength. With the development of disciplinary theory and computer simulation technology, complex mechanical product typically requires extensive use of simulation-based design and analysis tools, Despite the steady and continuing growth of computing power and speed, the computational cost of complex high-fidelity engineering analyses and simulations maintains pace. The high computational expense limits, or often prohibits, the use of such codes in engineering design and multidisciplinary design optimization (MDO). Meanwhile, the manufacturing industry competitively aims at shortening the product development and manufacturing cycles and reducing product development costs. Therefore, the conflicts between computational Accuracy and efficiency are is an important issue for engineering design of complex products.Metamodeling techniques are widely used in engineering design to address these concerns. The basic approach is to construct approximations of the analysis codes that are more efficient to run, and yield insight into the functional relationship between design variables and response. In this work, we investigate support vector regression (SVR) as a metamodel for approximating complex engineering analyses, and explores the basic theory and the key implementation technologies on metamodel based on support vector regression for engineering optimization problems. The dissertation carried out researches on the following topics and obtained the corresponding results.1) By comparing the advantages and disadvanteges of existing kinds of popolar metamodel methodology, SVR metamodel method was proposed. By using testing functions and engineering example to make comparative research on the precision of approximate models, results show SVR metamodel method is high efficiency and precision.2) Aiming at the optimization design problem with implicit objective performance functions, a design optimization method based on SVR metamodel and genetic algorithm (GA) is proposed, a framework based on the SVR and particle swarm optimization (PSO) for structure optimization design, and a multiobjective design optimization method based on SVR metamodel and improved Non-dominated Sorting Genetic Algorithm (NSGA-II) is proposed. The structure optimization of a microwave power divider is adopted as an example to illustrate the effectiveness of these design methods.3) Aiming at the robust optimization with uncertainty design problem of computationally intensive simulation models, a reduced approximation model technique based on SVR is introduced in order to improve the accuracy of metamodel. A framework based on SVR and GA is presented for robust optimization problems. The performances of SVR were compared with other existing metamodels under uncertainty. The applicability of the method is demonstrated using a two-bar structure system study, the results showed that the prediction accuracy of SVR model was higher than those of others metamodels, and the proposed optimization methodology is found to be accurate and efficient for robust optimization.4) Exiting MDO methods are reviewed, and the advantages and disadvantages of these methods are discussed and analyzed. Collaborative optimization (CO) is systematically investigated. In order to deal with complicated MDO problem, a novel MDO CO method based on SVR metamodel (SVR-CO) is proposed. By using typical coupled optimization example to make comparative research of three methods including SVR-CO, CO and Multidisciplinary Feasibility Method (MDF), SVR-CO is proven to be efficient and effective.