Refine
Is part of the Bibliography
- no (1)
Has Fulltext
- yes (1)
Year of publication
- 2021 (1)
Document Type
- Master's Thesis (1)
Institute
Language
- English (1)
Analysis of numerical crash simulation data using dimensionality reduction and machine learning
(2021)
Automakers find it challenging to analyze multiple computer-based numerical crash simulation data, with an increase in computer-based simulations. This downside causes the necessity to use data science techniques to automate the analysis of an ensemble of crash simulation data. This study aims at setting up a workflow to dimensionally reduce the simulation data, cluster it based on a behavioral pattern and analyze a particular behavioral pattern to obtain if-else rules to avoid or achieve that particular behavioral pattern. In this context, a behavioral pattern refers to an observed characteristic result that occurs because of specific input parameter values.
In this thesis, dimensionality reduction is undertaken using feature extraction algorithms, the knowledge behind the simulations is extracted and clustered based on an individual simulation’s behavioral pattern using unsupervised clustering algorithms, and the rules to avoid or achieve a particular behavioral pattern are extracted using decision tree algorithm and associative rule mining algorithm.
The workflow is applied to a simple side pole impact test, where a pole is impacted to an assembly of three parallel vertical members welded to a horizontal member. Using the workflow, every simulation data is represented using 50 lower-dimensional parameters and the simulations in the dataset are clustered into three different clusters based on three different behavioral patterns observed. One of the behavioral patterns observed is lower buckling of a vertical member, which is not a desired crash behavior to be observed in a crash energy absorption structure and this behavior should be avoided. The rules to avoid this behavioral pattern is further obtained using rule extraction.
The results from the study suggest that the workflow and the algorithms used in this thesis to analyze an ensemble of simulation data helps an engineer in efficiently representing simulation data using lower-dimensional embedding, extracting the underlying knowledge, and obtaining rules to avoid or achieve a certain behavioral pattern. Furthermore, there are many interesting directions worth perusing based on this thesis, (1) Achieving generalizability while using simulations of different mesh criteria. (2) Obtaining validation metrics to validate the results. (3) Using different datasets to evaluate and benchmark the approach given in this study. (4) Defining hyperparameter optimization strategies to efficiently optimize the hyperparameters of the machine learning algorithms used in this thesis.