Comparative Study of TCN+GAN and LSTM+CNN Architectures for Time-History Curves Prediction

Predicting the unobserved rear half of an acceleration time-history curve can make automotive safety evaluation more efficient. This study compares two deep-learning models, LSTM+CNN and TCN+GAN, for reconstructing post-impact acceleration signals from their observed front halves. The workflow includes signal preprocessing, statistical feature augmentation, and model training with reconstruction losses; TCN+GAN also uses an adversarial loss. We assess pointwise accuracy, temporal alignment, and physical consistency, and examine how the predictions affect downstream Head Injury Criterion (HIC) estimates. In the reported tests, TCN+GAN preserved key dynamic features more closely and produced more physically plausible signals than LSTM+CNN. The results suggest that it could serve as a surrogate in some simulation tasks and support injury prediction.