A domain-independent expert system for complex mechanical designs

In general, mechanical design is the process of creating concepts and specifying detailed physical systems and material components to achieve design goals. It is a search process for refinement from a large space of possibilities. Conventionally, the search process is mathematically modeled into constrained optimization problems. Many sophisticated numerical algorithms in nonlinear programming have been developed for solving the problem. However, many special features of practical mechanical designs have not been fully incorporated in the development of the solvers. The non-differentiability, the qualitative evaluation, and the designer's preference are examples which are difficult to be handled numerically. This inadequacy prompts the problem to be reexamined in the domain of modern Artificial Intelligence Techniques. However, available general purpose AI systems solve only simple design problems and are generally focused on narrow specified domains. This dissertation synthesizes the state of the art technologies in Artificial Intelligence, Statistics, Numerical Computation, and Fuzzy Set Theory, and present a hybrid general purpose Expert System to solve mechanical design problems. The new system is composed of a statistical learning model and a fuzzy redesign model. The statistical learning model derives domain specific knowledge by memorizing design samples. The "design experience" is correlated to predict initial designs when new design conditions are described. The fuzzy redesign solver iteratively improves any initial design based on qualitative judgments. The dependencies between design variables are automatically derived. Redesign decisions are made depending on the dependency, priority, and preference. The new approach has been verified and demonstrated using three reasonably complex practical design problems. They are the two bar truss design, welded neck flange design, and fixed tubesheet heat exchanger thermal differential design. The method proposed in this dissertation performed very well in automating poorly understood complex mechanical design processes proving its effectiveness, reliability, efficiency, and versatility.