Variance-based global sensitivity analysis for rear-end crash investigation using deep learning.

Area of Science:

• Transportation Safety
• Data Science
• Machine Learning

Background:

• Traffic accident injury severity (INJ-S) analysis is challenging due to rare events and data instability.
• Classical statistical models struggle with INJ-S prediction and factor prioritization.
• Rear-end collisions, the most frequent accident type, require deeper investigation into injury severity factors.

Purpose of the Study:

• To develop a robust methodology for analyzing traffic accident injury severity (INJ-S).
• To accurately predict INJ-S levels and identify significant contributing factors in rear-end accidents.
• To overcome limitations of traditional models in handling complex accident data.

Main Methods:

• A deep learning paradigm utilizing deep residual neural networks (DRNNs) with residual shortcuts.
• Integration of Variance-Based Global Sensitivity Analysis (VB/GSA) for factor impact assessment.
• Monte Carlo simulation applied to a DRNN model trained on North Carolina rear-end accident data (2010-2017).

Main Results:

• The developed methodology achieved an 83% overall accuracy in predicting INJ-S.
• The deep learning approach demonstrated superior performance compared to the ordered logistic regression (OLR) model.
• VB/GSA effectively identified the most significant attributes influencing rear-end crash injury severity.

Conclusions:

• The proposed deep learning and VB/GSA methodology offers a powerful tool for analyzing traffic accident injury severity.
• This approach provides accurate predictions and interpretable insights into factors driving accident outcomes.
• The findings can inform targeted safety interventions to reduce injury severity in rear-end collisions.