Two-stage physics-guided deep learning framework for impact force identification and source localization using a single sensor

doi:10.1016/j.ymssp.2025.113586

Two-stage physics-guided deep learning framework for impact force identification and source localization using a single sensor

2025-11-06

0

PRE

AI

DOI：10.1016/j.ymssp.2025.113586

Original

Original request for help

Save

Abstract

En 中文

Accurate identification of time-varying impact forces from noisy structural responses is critical for ensuring the operational safety of composite structures. Conventional physics-driven methods typically suffer from high computational complexity and sensitivity to system identification errors, whereas purely data-driven methods lack sufficient physical interpretability. To address these challenges, this paper proposes a novel two-stage physics-data hybrid-driven deep learning framework, termed the Physics-guided Impulse Response Dictionary Method (PG-IRDM). In the first stage, a Physics-guided Impulse Response Temporal Convolutional Network (PG-IRTCN) is constructed by integrating L1 sparsity constraints, frequency-domain consistency, and Tikhonov regularization to effectively predict impulse response functions and reconstruct impact forces across the structural surface. In the second stage, an expanded multi-sample dictionary built from impulse response similarities enables accurate impact localization with a single sensor, overcoming the limitations of conventional “localization-then-inversion” paradigms. Experimental results demonstrate that PG-IRDM achieves peak force prediction accuracies exceeding 95.41 % on both a composite flat plate and a foam-core blade, surpassing GTCN, DPH-TrLSTM, and NSC-Net by 6.93 %, 12.32 %, and 35.67 %, respectively, with impact timing errors below 1.2602 ms. Under sufficient data conditions, PG-IRDM also successfully accomplishes localization tasks that remain challenging for other methods, achieving a localization accuracy exceeding 91.57 %. Moreover, systematic sensor-position analysis indicates that central arrangement contributes to accurate force reconstruction, and edge arrangement enhances localization performance. Overall, PG-IRDM demonstrates excellent physical interpretability and superior noise resistance, offering a practical and scalable solution for intelligent monitoring and structural integrity assessment in complex composite systems.

Two-stage physics-guided deep learning framework for impact force identification and source localization using a single sensor

Abstract

Journal

Organization

Cited Papers