Work

Physics-informed deep learning for few-shot damage localization in carbon fiber reinforced polymer composites under thermal variations

Year: 2026

Type: article

Abstract: Guided wave (GW)-based damage localization in carbon fiber reinforced polymer (CFRP) composites is fundamentally challenged by thermal variations and sparse sensor configurations, which induce severe baseline distortions and imaging artifacts. Traditional physical algorithms suffer from geometric ambiguity, while purely data-driven artificial intelligence (AI) models are hindered by opaque interpretability and prohibitive data dependency. This paper proposes a physics-informed deep learning (PIDL) framework, an artificial-intelligence-driven approach, for few-shot damage localization under thermal variations. The network is architected with a dual-stream input design, comprising a data-driven layer enhanced by a one-dimensional coordinate attention (1D-CA) mechanism and a physics-informed layer built upon a differentiable time-of-flight (TOF) physical imaging layer. A hybrid physics-constrained loss function is formulated to suppress non-physical artifacts and enforce physical consistency. GW signals are experimentally collected from manufactured CFRP composite plates, and a comparative investigation is performed to thoroughly assess the PIDL performance. Results indicate that:(1) The presence of physical penalties effectively mitigates artifacts and achieves robust localization across thermal gradients; (2) The model demonstrates superior spatial generalization that the integration of acoustic physical constraints acts as a crucial regularization mechanism, maintaining a highly accurate mean relative error (MRE) of 3.88% on completely unseen damage coordinates; (3) By implementing a few-shot transfer learning strategy, the PIDL model shows adaptability for dynamic operational environments (45 degrees Celsius (°C)-75 degrees Celsius (°C)). This work offers a physically interpretable and data-efficient paradigm for structural health monitoring (SHM) in dynamic operational environments.

Source: Engineering Applications of Artificial Intelligence

DOI: 10.1016/j.engappai.2026.116569