Process design optimization based on metamodeling and metaheuristic techniques

Abstract

This dissertation explores the use of mathematical and statistical tools to analyze, control and optimize manufacturing processes. Table 8.1 summarizes the topics analyzed and the contributions of this dissertation. The main topic discussed in this thesis is related to the features that have to be taken into account to select, according to the analyzed process: 1. the metamodel technique; 2. the sampling strategy; 3. the optimization algorithm All these problems were analyzed from two di erent point of views. Attacking the problems from a Computer Science angle has led to the development of a general version of the methodologies. In contrast, it is also crucial to analyze the process from a mechanical point of view trying to detect peculiarities that may simplify the computational e ort to solve the problem making use of the available knowledge. Chapter 2 provides a brief introduction of the most used approaches to de ne input-output relationships. It is pointed out that each technique is a ected by many limitations that could signi cantly a ect the accuracy of the model under certain conditions.Table 8.1: Dissertation Contributions. Chapter Theory Applications Chapter 1 Introduction, research statement and scope of the thesis Chapter 2 Machine learning techniques introduction Impression die forging Chapter 3 Heuristic technique to optimize neural network performance Extrusion, rolling and shearing Chapter 4 Kriging metamodel for mixed continuous/discrete problems Incremental sheet forming: thickness distributon Chapter 5 Manufacturing processes problem modeling Incremental sheet forming: temperature prediction Chapter 6 Adaptive KPI prediction based on response surface projection through similarity function Remote laser welding Chapter 7 Multi Objective Techniques. Development of a GDE3 based algorithm. Porthole extrusion Chapter 3 discusses the use of heuristic algorithms to solve these limitations. Di erent techniques were developed to improve the performance of a neural network metamodel. In particular three heuristics were developed and used to: 1. select the network architecture; 2. select the starting weights; 3. reduce the training time through an hybrid algorithm (simulated annealing+ backpropagation) Chapter 4 presents a novel kriging metamodel to solve problem characterized by both continuous and discrete variables. The model was coupled with a customized sampling strategy to reduce the number of experiments to reach a required accuracy. Typically, a speci c DoE method is most suitable in combination with each individual metamodel formulation. The proposed designs try to maximize a space- lling property. This feature assures a balanced predictive performance of the approximation model throughout the investigated model space. To collect training data e ciently, the locations for sampling points have to be chosen systematically thus assuring a maximum gain in information with minimal e ort. In particular the space lling criterion was considered only around a feasible region denoted as process window. Chapter 5 discuss the use of customized model based on prior knowledge of the process. According to that chapter 6 presents a new methodology with which a metamodel is developed making use of qualitative knowledge and/or historical data on similar problems. The metamodel try to iteratively develop new response surfaces through a geometrical projection based on a similarity function. Finally chapter 7 discuss the use of evolutionary algorithm to solve multi objective problems making use of the previous developed metamodel. To conclude, it is believed that this dissertation explores machine learning and optimization techniques for manufacturing from many angles, and that several of the ideas presented here will be useful both in practice and for theoretic studies