Xianping Du

Date of Award

Fall 2019

Access Type

Dissertation - Open Access

Degree Name

Doctor of Philosophy in Mechanical Engineering


Mechanical Engineering

Committee Chair

Feng Zhu, Ph.D.

First Committee Member

Fady Barsoum, Ph.D.

Second Committee Member

Sathya Gangadharan, Ph.D.

Third Committee Member

Hong Liu, Ph.D.

Fourth Committee Member

Hongyi Xu, Ph.D.


This study develops a systematic design methodology based on data mining theory for decision-making in the development of crashworthy vehicles. The new data mining methodology allows the exploration of a large crash simulation dataset to discover the underlying relationships among vehicle crash responses and design variables at multiple levels and to derive design rules based on the whole-vehicle safety requirements to make decisions about component-level and subcomponent-level design. The method can resolve a major issue with existing design approaches related to vehicle crashworthiness: that is, limited abilities to explore information from large datasets, which may hamper decision-making in the design processes.

At the component level, two structural design approaches were implemented for detailed component design with the data mining method: namely, a dimension-based approach and a node-based approach to handle structures with regular and irregular shapes, respectively. These two approaches were used to design a thin-walled vehicular structure, the S-shaped beam, against crash loading. A large number of design alternatives were created, and their responses under loading were evaluated by finite element simulations. The design variables and computed responses formed a large design dataset. This dataset was then mined to build a decision tree. Based on the decision tree, the interrelationships among the design parameters were revealed, and design rules were generated to produce a set of good designs. After the data mining, the critical design parameters were identified and the design space was reduced, which can simplify the design process.

To partially replace the expensive finite element simulations, a surrogate model was used to model the relationships between design variables and response. Four machine learning algorithms, which can be used for surrogate model development, were compared. Based on the results, Gaussian process regression was determined to be the most suitable technique in the present scenario, and an optimization process was developed to tune the algorithm’s hyperparameters, which govern the model structure and training process.

To account for engineering uncertainty in the data mining method, a new decision tree for uncertain data was proposed based on the joint probability in uncertain spaces, and it was implemented to again design the S-beam structure. The findings show that the new decision tree can produce effective decision-making rules for engineering design under uncertainty.

To evaluate the new approaches developed in this work, a comprehensive case study was conducted by designing a vehicle system against the frontal crash. A publicly available vehicle model was simplified and validated. Using the newly developed approaches, new component designs in this vehicle were generated and integrated back into the vehicle model so their crash behavior could be simulated. Based on the simulation results, one can conclude that the designs with the new method can outperform the original design in terms of measures of mass, intrusion and peak acceleration. Therefore, the performance of the new design methodology has been confirmed.

The current study demonstrates that the new data mining method can be used in vehicle crashworthiness design, and it has the potential to be applied to other complex engineering systems with a large amount of design data.