Evidence-Guided Attention Neural Network for Structural Crack Identification with Multi-Source Sensors

The integration of complementary sensing modalities provides a basis for accurate crack identification in advanced aircraft structures. In this context, PZT transducers are sensitive to incipient damage through guided-wave interrogation, whereas strain gauges capture quasi-static deformation. Prevailing fusion paradigms, however, encounter an interpretability–adaptability dilemma. Model-based approaches lack robustness to sensor degradation, while data-driven attention methods sacrifice transparency. To resolve this trade-off, an Evidence-guided Attention Neural Network (EANN) is proposed. Its central methodological contribution lies in repositioning Dempster–Shafer (D-S) evidence theory from a terminal fusion operator to an upstream credibility feature extraction module. Evidence-derived credibility features, comprising belief entropy, inter-source similarity, and Kalman-filtered residuals, drive the attention weight optimization and endow the learned channel weights with physically interpretable evidential meaning. Ablation experiments confirm that the observed improvement arises from the interaction between the evidence-guided credibility representation and adaptive attention weighting, with neither component sufficient on its own. The framework fuses quasi-static strain measurements with active piezoelectric guided-wave interrogation, which offers high sensitivity to incipient damage but remains vulnerable to channel degradation. Experiments on aluminum tensile plates and trapezoidal wing skin specimens show that EANN maintains identification accuracy under simulated sensor anomalies and partial failures by attenuating compromised channels without explicit fault detection, providing an uncertainty-aware fusion framework for online structural health monitoring of aerospace structures.