Predicting occupant response curves in vehicle crashes via Attention-enhanced multimodal temporal Network.
Journal:
Accident; analysis and prevention
Published Date:
Jun 17, 2025
Abstract
Accurately predicting safety responses, especially occupant crash response curves across multiple body regions, plays a crucial role in advancing vehicle crash safety by enabling design optimization and reducing the reliance on costly physical testing and simulations. Machine learning methods have demonstrated good performance in this field, but existing approaches often face challenges in integrating multimodal data and handling multi-task temporal predictions. To address these issues, this work proposes a novel Attention-enhanced Multimodal Temporal Network (AMTN) for predicting occupant crash response curves across multiple body regions during crashes. AMTN integrates numerical parameters and vehicle body crash pulses through a feature extraction module, organically fuses multimodal features via cross-attention mechanisms, and decodes shared features using a modified Temporal Convolutional Network (TCN) with local sliding self-attention. A dynamic adaptive loss and multiple output layers are utilized to evaluate the importance of each task, iteratively update the learning priorities, and finally achieve multi-task prediction. Experiments on the engineering-obtained data demonstrate that the crash response curves predicted by AMTN achieves an average ISO (International Organization of Standards) rating of 0.835 across 11 crash response curves, with critical regions exceeding 0.9. In engineering applications, an ISO rating greater than 0.8 indicates that the predicted curve closely matches the reference curve. Therefore, the experimental results demonstrate that the proposed method has effectively learned the characteristics of the training data and is capable of producing accurate predictions. In summary, this work advances multimodal deep learning for crash safety by enabling efficient, accurate, and interpretable multi-task curve predictions. Consequently, it can be applied to data-driven vehicle safety development, offering significant engineering value by enhancing both development efficiency and quality.
