Accelerating Bayesian inference of structural damage detection based on physics-guided CNN and zero-shot transfer learning

Abstract

Bayesian inference for structural damage identification faces prohibitive computational demands due to high-dimensional parameter spaces and repeated evaluations of costly finite element (FE) models during Markov chain Monte Carlo (MCMC) sampling. This study introduces an accelerated Bayesian framework integrating a physics-guided convolutional neural network (PhyCNN) with zero-shot transfer learning to overcome these limitations. The PhyCNN surrogate model embeds fundamental mechanics principles via physics-constrained loss functions, replacing FE simulations to rapidly predict modal parameters for damage states. To address domain discrepancies between simulation (source) and operational (target) environments, an adversarial sparse autoencoder establishes domain-invariant latent representations through sparse feature alignment, enabling training-free adaptation to target tasks. Validated on an experimental three-story frame and a numerical cable-stayed bridge model, the proposed framework achieves an order-of-magnitude computational acceleration over conventional MCMC while maintaining high accuracy. Ablation studies demonstrate that integrating physical constraints reduces damage localization errors by 5.14% compared to purely data-driven models. The methodology’s generality is underscored by its compatibility with diverse sampling-based uncertainty quantification methods and extensibility to systems with computationally intensive likelihoods. By synergizing physics-aware deep learning with cross-domain transfer, this work bridges the gap between simulation-driven training and real-world structural health monitoring, offering a practical pathway for real-time Bayesian diagnostics in large-scale infrastructure.