Impact force identification via a physics-informed sparse coding network

doi:10.1016/j.cja.2026.104362

Impact force identification via a physics-informed sparse coding network

2026-07-08

0

OA

AI

DOI：10.1016/j.cja.2026.104362

Original

Original request for help

Save

Abstract

En 中文

Accurate identification of impact forces is essential for Structural Health Monitoring (SHM) and safety assessment of aerospace composite structures. Impacts from foreign objects may induce barely visible damage, posing risks to flight safety. While traditional regularization-based inverse methods are sensitive to parameter selection, deep learning approaches often depend on large labeled datasets. This paper proposes a Physics-Informed Sparse Coding Network (PISC-Net) for impact force identification. By exploiting the intrinsic sparsity of impact forces, a non-convex sparse regularization model is formulated based on the Minimax Concave Penalty (MCP). The model is solved via a Proximal Gradient Descent (PGD) algorithm, which is then unrolled into a sparse coding network to enable adaptive parameter tuning and interpretable design. To incorporate physical knowledge, the system transfer matrix is embedded as a non-trainable decoding layer. The training follows a self-supervised strategy by minimizing the reconstruction error between predicted and measured response signals, eliminating the need for ground-truth labels. Simulation and experimental results demonstrate that PISC-Net achieves superior accuracy and robustness compared to Tikhonov regularization, underscoring its potential for practical impact monitoring in advanced aerospace structures.

Keywords:

Impact force identification

Physics-informed neural network

Algorithm unrolling

Self-supervised learning

Sparse coding

Aerospace structural health monitoring

Impact force identification via a physics-informed sparse coding network

Abstract

Journal

Organization