A data mining method for structure design with uncertainty in design variables

Abstract

The traditional structural optimal design methods aiming to generate a global optimum may fall into the unfeasible domain due to the presence of uncertainty. This issue can be addressed by generating a group of satisfactory design or sub-design regions rather than a single optimal one. A data mining method has been recently developed based on the decision tree technique and applied to the engineering structural design by learning from a big design dataset. It solves the design problems in an explainable way and helps designers understand design problems efficiently. This method, based on the traditional decision tree algorithm, however, cannot handle uncertain data. In this work, a new decision tree for uncertain data (DTUD) method is developed based on the joint probability distribution of design variables for the engineering design. Its high accuracy is verified by comparing it with the traditional decision tree using nine datasets selected from a publicly available repository. To demonstrate the performance of this method in structural design problems, it is implemented in the design of a thin-walled energy-absorbing structure subjected to crash loading. With assumed probability distribution on the uncertain data, an uncertain decision tree is built, which generates designs with expected performance effectively and efficiently. Besides, the deterioration of design performance due to uncertainty can be captured by the new decision tree. This further helps improve the reliability of the new designs.