Artificial neural network modeling for mechanical design expert systems: Elastomeric vibration isolator case study

An artificial neural network based methodology for mechanical component functional modeling is presented that can be used as the kernel for expert design systems. The proposed methodology captures several generations and variations of design data within a general product type. The artificial neural network methodology is improved with a demonstrated process for selecting the best mapping networks and tests for validating the primary and secondary relationships between the training parameters and the output parameters. This generalization process is also shown to be successful in finding spurious data points. Mapping parameter reduction concepts are developed for minimizing the “curse of dimensionality” without losing the capability of capturing the highly nonlinear, interdependent relationships.; The network training process is further improved with demonstrated methods for supplementing the training data with heuristics, limiting conditions, and analytical based training data. With this supplementation it is shown that the mapped domain can be significantly expanded beyond the base training domain. A data transformation technique is developed that significantly improves the accuracy of the mapping function for the large training domain range coupled with a large variation in training data density. Also proposed and developed is the use of uncertainty for both the training data and the mapped results. These improvements to the traditional artificial neural network mapping process result in a robust expert design system kernel.; These mapping process concepts are validated on a product line of elastomeric vibration isolators spanning several engineering generations. These isolators embody many design variables with multiple design parameters. This database is representative of many in the mechanical design field that can be leveraged for new designs. This induction process reduces the dependence of traditional expert system development in capturing the basic and heuristic knowledge from several generations of human experts.; The final product of this research is a robust expert design system kernel for elastomeric vibration isolators that has induced it's mapping function from real data that contains all the relationships difficult to capture by traditional modeling techniques. The final kernel is significantly more generalized than the specific design points used to train it.