Exploring Possibilities of AI for Generating Synthetic Data of Highly Dynamic Processes to Improve ML Model Generalization
Open Access
- 22 September 2026
- conference paper
- Published by MDPI AG in Engineering Proceedings
- Vol. 157 (1) , 6
Abstract
This paper investigates AI models for generating synthetic data from highly dynamic and size-limited lateral-sled-test measurements for the purpose of training data augmentation. Given that lateral sled tests serve as a cost-efficient proxy for full-scale crash experiments while covering only a limited subset of deceleration, intrusion, and restraint-system configurations, the problem is formulated as the conditional generation of multivariate transient response signals under varying boundary conditions. The study benchmarks modern generative models for multivariate time-series synthesis with the future objective of using synthetic training data augmentation to improve robustness of data-driven models on previously unseen test configurations. Rather than treating this augmentation as a simple increase in sample size, the generated data are assessed along three complementary dimensions: temporal and distributional fidelity, diversity and coverage of underrepresented operating regimes, and physical plausibility of the synthesized responses.Keywords
Funding Information
- Programme Johannes Amos Comenius (CZ.02.01.01/00/23_020/0008528)
This publication has 15 references indexed in Scilit:
- Generative Artificial Intelligence Models: A SurveyArtificial Intelligence Review, 2026
- WaveStitch: Flexible and Fast Conditional Time Series Generation With Diffusion ModelsProceedings of the ACM on Management of Data, 2025
- Physics-consistent machine learning with output projection onto physical manifoldsCommunications Physics, 2025
- MetaIndux-TS: Frequency-Aware AIGC Foundation Model for Industrial Time SeriesIEEE Transactions on Neural Networks and Learning Systems, 2025
- Facing & mitigating common challenges when working with real-world data: The Data Learning ParadigmJournal of Computational Science, 2025
- Synthesizing Distribution Grid Congestion Data Using Multivariate Conditional Time Series Generative Adversarial Networks2024
- Diff-MTS: Temporal-Augmented Conditional Diffusion-Based AIGC for Industrial Time Series Toward the Large Model EraIEEE Transactions on Cybernetics, 2024
- A synergistic fusion of shallow and deep generative model to enhance machine learning efficacy and classification performance in data-scarce environmentsInternational Journal of Information Technology, 2024
- A comprehensive survey for generative data augmentationNeurocomputing, 2024
- Sig‐Wasserstein GANs for conditional time series generationMathematical Finance, 2023