Research On Ensemble Of Surrogates And Sequential Sampling Strategies For Product Design Optimization

Complex mechanical product usually consists of lots of components which couple and act each other, complicating the whole design optimization process. As one of the effective tools to solve such problems, computer simulation technique is applied to simulate the physical experiments based on the digital models of object problem, and the data obtained from these simulations can be used to achieve the optimal solution through optimization algorithm. However, with the quality requirements of the mechanical products increasing, the accuracy of simulation models becomes higher and higher and the computational expense is exponentially increased at the same time, dramatically hampering the application of this technique in industry. In order to overcome this challenge, the Surrogate Based Design Optimization (SBDO) methods are proposed by many researchers. Using surrogate models instead of complicated time-consuming computer simulations and with the aid of proper optimization algorithm, SBDO methods can achieve the goals of reducing computation burden and alleviating problem complexity, which show great potential in engineering applications.As a SBDO method, ensemble of surrogates and sequential sampling strategies based product design optimization provides an effective approach for product design. It takes full advantages of highly accurate and robust ensemble models, sequential sampling methods for rational configuration of data sets and system decomposition strategy conforming to engineers'design habits. In this method, ensemble models are used to replace computational expensive implicit objective functions and constraints as the fundamental of optimization process. Then sequential sampling methods are adopted to ensure the accuracy of ensemble models. At last, the coordination and decomposition strategies are constructed which provide an effective and flexible way for the solution of product design optimization problem. Thus, ensemble of surrogates, sequential sampling methods along with the coordinated decomposition strategy constitute the key roles of ensemble of surrogates and sequential sampling strategies based product design optimization method. In this thesis, further researches on these three key issues are carried out.Firstly, the mathematical model, the related concepts and terminology of ensemble of surrogates and sequential sampling strategies based multidisciplinary design optimization problem are given. The deficiencies of existing researches are illustrated later. And then, research framework of the key techniques is presented.Secondly, the Ensemble of Surrogates with Hybrid methods using Global and Local measure (ES-HGL) is proposed. According to the location of prediction points, the weight factors of surrogate model are calculated by global measure or local measure. Then the weight factors are corrected by hybrid weight factor according to the defects of each measure. The weight factors computed from both measures are based on the same error matrix. ES-HGL method has almost the same computational cost compared with each single measure, which provides strong support for the following optimization process.Next, Voronoi diagram and Adaptive Sampling Radius based sequential sampling method (ASR-Voronoi) is proposed. It can limit the location of new sample points within the sensitive Voronoi cell, new sample points are always added in areas with large modeling errors, thus to achieve a stable sequential sampling process. Considering the defects of existed methods, ASR-Voronoi introduces the concept of adaptive sampling radius to measure the difficulty of fitting the design space of surrogate models. The configuration of sample points is accomplished by the value of adaptive sampling radius, balancing the exploitation and exploration feature in sequential sampling method. ASR-Voronoi method is an effective surrogate model updating strategy, which can provide a highly accurate ensemble model for optimization process.And then, Analytical Target Cascading based on Combination Physical Programming (ATC-CPP) is proposed. This method attempts to improve the inefficiency of ATC-PP method that the consistency constraints must be strictly satisfied. In this case, the combination physical programming is introduced into ATC framework, where the preference function is simplified and the aggregate objective function is corrected to satisfy the engineers'design habits.After that, the research theories mentioned above are integrated into the ensemble of surrogates and sequential sampling strategies based product design optimization framework. The achievements of this paper are successful applied to the design optimization of a super heavy CNC machine tool, showing the effectiveness of the three key techniques proposed in solving product design problems.Finally all the contributions in this thesis are summarized, and the future research directions are prospected.