Abstract
In the present study, the amount of fragments generated and their travel distances due to vehicle collision with concrete median barrier (CMB) was analyzed and predicted. In this regard, machine learning was applied to the results of numerical analysis, which were developed by comparing with field test. The numerical model was developed using smoothed particle hydrodynamics (SPH). SPH is a mesh-free method that can be used to predict the amount of fragments and their travel distances from concrete structures under impact loading. In addition, deep neural network (DNN) and gradient boosting machine (GBM) were also employed as machine learning methods. In this study, the results of DNN, GBM, and numerical analysis were then compared with the conducted field test. Such comparisons revealed that numerical analysis generated lower error than both DNN and GBM. When prediction results of both the amount of fragments and their travel distances were considered, the result of DNN showed smaller errors than that of GBM. Therefore, in studies where machine learning is used to predict the amount of fragments and their travel distances, careful selection of an appropriate method from the various available machine learning methods such as DNN, GBM, and random forest is absolutely important.
Highlights
In Korea, to ensure passenger protection, the structural performance of concrete median barriers on expressways are evaluated periodically through certain tests [1]
mean absolute error (MAE) was smaller and R2 was larger in gradient boosting machine (GBM) than in deep neural network (DNN)
The error between the prediction and the analysis results was smaller in DNN, while performance of the regression model was better in GBM
Summary
In Korea, to ensure passenger protection, the structural performance of concrete median barriers on expressways are evaluated periodically through certain tests [1]. In the structural performance test, the weight and travel distance of the fragments of concrete median barriers should not exceed 2 kg and 2 m, respectively. There has been no parametric study reported to-date on the prediction of concrete fragments and their travel distance. In the present study, we developed an analytical model using SPH to predict structural performance of median barriers, which depends upon certain major variables viz. Using the analytical results as training data for machine learning, it was possible to accurately predict the amount of fragments generated after a collision and their travel distances. The second step was construction of DNN and LightGBM to predict concrete fragments and their travel distance based on the developed numerical model (Sections 4.1–4.4). Administration (FHWA) of the United States to analyze collisions between vehicles and roadside facilities
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
