Neural extensions to robust parameter design

Robust parameter design (RPD) is implemented in systems in which a user wants to minimize the variance of a system response caused by uncontrollable factors while obtaining a consistent and reliable system response over time. Typically, quadratic regression is deemed sufficient to specify a process model of model system behavior. We propose the use of artificial neural networks (ANNs) to compensate for highly non-linear problems that quadratic regression fails to accurately model.;RPD is conducted under the assumption that the relationship between the system response and controllable and uncontrollable variables does not change over time. Since degradation in the system response will almost certainly occur; this assumption will inevitably be violated. We propose a methodology to find a new set of settings that will be robust to moderate system degradation while remaining robust to noise variables within the system. An algorithm is presented for this enhanced RPD analysis utilizing both quadratic regression and two specific artificial neural network architectures.;RPD has been well developed on single response problems. Sparse literature exists on dealing with multiple responses in RPD and most methods utilize a subjective weighting scheme. To account for multiple responses, we examine the use of factor analysis on the response data. Linear combination techniques are also developed in the case that more than a single factor is retained in the analysis.;All the proposed techniques are applied to textbook applications to demonstrate their utility. An Air Force application problem is then examined to demonstrate the new technique's potential on a real-world problem that is highly non-linear. The application is a detector developed to detect anomalies within hyper-spectral imagery.;The results of this research include successful implementation of artificial neural networks in RPD. These artificial neural networks can be utilized when faced with a highly non-linear problem. Also, new settings are developed that are shown to be superior to traditional robust settings when a system is subject to performance degradation. A new methodology of approaching multiple response problems is developed which shows promise. Finally, the anomaly detector is further enhanced through the use of artificial neural networks to determine robust settings and alternate settings when degradation is expected.