Physics-Guided Self-Attention for Metallic Plate Impact Localization With FBG Under Low-Sampling-Rate Constraints

Abstract

This study presents a physics-informed self-attention neural network for real-time impact localization in metallic structures using Fiber Bragg Grating sensor networks, overcoming the critical Nyquist-rate limitation of conventional time-difference-of-arrival (TDOA) methods. The proposed framework achieves centimeter-level accuracy under low-sampling-rate. Two key innovations enable this breakthrough: 1) physics-guided feature extraction that explicitly models Lamb wave dispersion and attenuation characteristics, 2) optimized single-head attention mechanism for efficient spatiotemporal correlation modeling. Experimental validation on aerospace-grade titanium plates demonstrates robust performance across 120 impact events with a mean absolute error of 1.18cm at 1 kHz sampling. This work bridges the gap between high-accuracy localization and low-speed interrogation systems, with direct applicability to spacecraft structural health monitoring, where power and bandwidth limitations prohibit high-speed data acquisition. Future directions include multi-impact resolution and on-orbit environmental adaptation.