| Popis: |
Sensitivity analysis is used in complex design by assisting a decision-maker to identify the most important parameters to control and manage during the design process. This sensitivity analysis can achieved by quantifying the sources of observed variance in the system outputs to be optimized by the decision-maker. Most sensitivity analysis methods are based on using a linear model to quantify the effect of each input parameter on the system output. This approach can fail to recognize parameters that have a strong but non-linear effect on the output. This research investigates the use of kriging models to quantify both linear effects and spatial effects for each of the potential model parameters. The result is an improved ability to identify and quantify the effect of input factors when the output is nonlinear and non-monotonic. The design of most modern systems is complex. The systems being designed are complex and the process to design them is complex. The systems are complex due to a potentially large number of parameters that must be managed and due to the interactions that may exist between these large number of parameters. The process of design is complex because of the need to communicate not only the large number of design parameters with members of a design team that may be distributed across the world and organizations but also the need to communicate the impact of changing design parameters have on the many different design elements (subsystems). Often these subsystems are designed by teams that are given targets for performance metrics and are not given additional details on how their subsystem impacts the overall system performance. The result is a design process that may require many iterations to converge, increasing the both the time and cost to deliver the system design. There are also situations in which it becomes nearly impossible for the design to converge while iterating. There is a need in design to better manage the information in the design process from the early stages of conceptual design though to the final stages of production. The conceptual design process is typically characterized by exploring many design options to evaluate their ability to most effectively achieve the desired system performance. The level of detail in this early stage of design is frequently quite low to allow the efficient exploration of a large tradespace. The level of detail is also kept low to permit easier comparison of design alternatives. Even with these simplifications during conceptual design the end result is a tradespace that may still be quite complex and difficult to explore and understand . It may also not be a very accurate representation of the possible final designs. As a result, the best use of this conceptual design process may be to remove design options that are obviously not going to achieve the desired system performance goals rather than specifying the single design that will be detailed. In this context, the conceptual design process can be performed through hypothesis testing to prove that a design (a set of design options or parameters that specify a design) performs worse than the other designs and should be excluded from the set of potential designs. The difficulty with performing hypothesis testing during conceptual design because of the large amount of uncertainty that frequently exists. Uncertainty in conceptual design can come from many sources. The first possible source is from using a mathematical model of physical reality. Reality is a function of nearly infinite factors. A mathematical approximation to reality must use a subset of these potential factors. One result is that a single parameter may be used to quantify the possible variability of a large number of parameters. An example of this maybe to treat the density of an alloy as a random variable rather than specify exactly the chemical make-up, manufacturing processes, and current temperature of the alloy. These additional parameters that aren’t included in the model are treated as ’noise’ sources. A second result of using mathematical relationships to map the input space to the output space is its inability to exactly quantify all of the possible relationships that may exist 1 . As a result, uncertainty can be introduced into design because of the form of the models and through parametric uncertainty. The approach often taken to reduce these types of uncertainty is to use higher fidelity models, i.e. more parameters and more complex relationships (typically are computationally expensive to evaluate). A second source of uncertainty in design that is seldom studied is the uncertainty in preferences of multiple performance levels. Often a design is initiated with a set of system performance goals or targets. Throughout the design process, the customer’s preferences may change for multiple reasons; 1) it may take an extended period of time |
