Leveraging hybrid knowledge-based and data-driven modeling techniques for complex contact/impact phenomena

Abstract

Accurate modeling of the contact/impact process between complex contacting surfaces is crucial for understanding and predicting dynamic behaviors of intricate systems. Although data-driven methods address certain limitations inherent in traditional physical models, their ${}^{\prime }$black-box${}^{\prime }$ nature limits the model’s interpretability and credibility. Moreover, existing approaches frequently struggle with challenges from sparse data, noise interference, and extrapolation tasks. To address these issues, this study proposes a dually knowledge-data driven strategy that integrates physical constraints with data-driven methods to enhance the interpretability, generalization and robustness of traditional contact/impact network models. Taking the interaction process between the inner bore and projectile as a case study, a finite element model of contact/impact process is firstly revisited to generate the sample dataset. Subsequently, effective physical information is extracted, and the collaborative mechanism between physical constraints and data-driven techniques is explored. The sample signals are subsequently subjected to pre-processing procedures including noise injection, data sparsification, and extrapolation treatment. These processed datasets then serve as the foundation for constructing knowledge-data driven models via back-propagation algorithms. Through comprehensive comparisons with purely data-driven models, the proposed hybrid modeling strategy demonstrates superior predictive performance across diverse scenarios, including noisy conditions, sparse datasets, and extrapolation tasks.