Assessment of the Modeling Abilities of Neural Networks

The treatment of manufacturing problems, whether in process control, process optimization, or system design and planning, can be helped by input-output models, namely, relationships between input and output variables. Artificial neural networks present an opportunity to "learn" empirically established relationships and apply them subsequently in order to solve a particular problem. In light of the increasing amount of applications of neural networks, the objective of this thesis is to evaluate the ability of neural networks to generate accurate models for manufacturing applications. Various neural network models has been tested on a number of "test bed" problems which represent the problems typically encountered in manufacturing processes and systems to assess the reliability of neural network models and to determine the efficacy of their modeling capabilities. The first type of problem tested on neural networks is the presence of noise in experimental data. A method to estimate the confidence intervals of neural network models has been developed to assess their reliability, and the proposed method has succeeded for a number of the models of the test problems in estimating the reliability of the neural network models, and greater accuracy may be achieved with higher-order calculations of confidence intervals which would entail increased computational burden and a higher requirement of precision for the parametric values of the neural network model. The second type of problem tested on neural networks is the high level of nonlinearity typically present in an input-output relationship due to the complex phenomena associated within the process or system. The relative efficacy of neural net modeling is evaluated by comparing results from the neural network models of the test bed problems with results from models generated by other common modeling methods: linear regression, the Group Method of Data Handling (GMDH), and the Multivariate Adaptive Regression Splines (MARS) method. The relative efficacy of neural networks has been concluded to be relatively equal to the empirical modeling methods of GMDH and MARS, but all these modeling methods are likely to give a more accurate model than linear regression. Thesis Supervisor: Professor George Chryssolouris Title: Associate Professor of Mechanical Engineering