A noise-robust hybrid neural network framework for damage localization in plates

Vibration-based structural health monitoring serves as a critical technology for ensuring the operational integrity of engineering systems. However, conventional methods relying solely on global natural frequency shifts fail to accurately locate damage in symmetric plate structures due to the non-uniqueness problem and high susceptibility to measurement noise. To address these limitations, this study proposes a triangulation-informed data-driven framework. The methodology introduces a hybrid input strategy that integrates multi-state frequency shifts—harnessing ambient temperature variations as a passive data augmentation technique—with spatial mode shape amplitudes extracted from three strategic Anchor Points, effectively resolving geometric ambiguities. Furthermore, a split-network architecture is developed to eliminate crosstalk interference. To ensure practical applicability, a simultaneous noise injection training strategy based on experimental modal analysis resolution limits (±0.10 Hz frequency and ±1% amplitude noise) is implemented. Numerical investigations on a clamped-free-clamped-free plate demonstrate that while standard frequency-based networks fail catastrophically under representative noise, the proposed hybrid framework successfully restrains the spatial deviation, ensuring reliable isolation of the damage zone. Under ideal conditions, the architecture achieves sub-element localization precision, offering a validated, scalable, and noise-resilient proof-of-concept for damage assessment in symmetric planar structures prior to experimental implementation.