Vehicle Occupant Kinematics prediction using Machine Learning: A study to understand applications of machine learning in prediction of vehicle occupant kinematics during crash and pre-crash
Examensarbete för masterexamen
Automotive engineering (MPAUT), MSc
Reddy, Venkata Sai Kumar
This work presents surrogate models using deep learning and statistical techniques for predictive modeling of occupant kinematics in crash scenarios. Finite Element Methods (FEM) involving discretization of Partial Differential Equations (PDEs) are widely used for vehicle crash and injury analysis involving Finite Element human body models (FE-HBMs). Doing a parametric study of pre-cash and in-crash scenarios involving a complex FE-HBM model depicting a detailed representation of the human anatomy is computationally expensive and time consuming, mainly due to the long pre-crash event. This thesis aims at evaluating machine learning models along with dimensionality reduction tools as the inexpensive surrogates for determining the kinematics of the vehicle occupants. A FE model of the Hybrid III crash test dummy was used to obtain nodal displacement outputs of the dummy for crash scenarios which in turn are the outputs predicted from the surrogates. These outputs from FE simulations form a large data set in high dimensional format. Performing machine learning with this large data set can become complex. This thesis utilizes one-dimensional convolutional autoencoder (AE-1D) and Principle component analysis (PCA) to obtain a compressed version of the original data. Supervised learning is used to train the surrogates on this low dimensional space with a set of crash parameters acting as inputs. Gaussian process regression (GPR), Feedforward neural networks (FFNN) and Random forest regression (RFR) were used as the machine learning models to build the surrogates. LS-Dyna FE solver was used to generate the sample results for a set of crash parameters. We use in-crash kinematics instead of pre-crash due to the computational limitations in simulating the required samples. For the same reason, a complex FE Human body model such as SAFER or THUMS model was not used. This thesis provides an understanding of using statistical and neural network based dimensionality reduction methods paired with different ML models on the accuracy of the kinematics predictions. The accuracy of the predictions varies considerably depending on the use of PCA or AE-1D for dimensionality reduction and also the sample size. PCA and AE-1D have similar performance in compressing and decompressing the original data. AE-1D paired with either FFNN or Gaussian process provided the most accurate predictions in most sample sizes.
FEM , Human body model , Hybrid III model , SAFER , Machine learning , Convolutional Autoencoder , Neural networks , Principal component analysis , Gaussian process , Surrogate modelling