A semi-supervised approach for structural damage state diagnosis under severely imbalanced class distributions

Rapid and accurate assessment of structural damage after earthquakes is crucial for improving emergency response efficiency and optimizing decision-making. Traditional nonlinear time-history analysis (NLTHA) is computationally expensive and relies on precise structural parameters, limiting its applicability for large-scale or rapid evaluations. While data-driven methods generally require a large amount of labeled data and are susceptible to class imbalance, resulting in insufficient recognition of minority classes, especially severe damage. To address the challenges of scarce labeled data and highly imbalanced classes in practical scenarios, this paper proposes a semi-supervised structural damage diagnosis model that integrates deformable convolutional residual networks (DCRN) with an imbalance-aware mechanism. This method uses continuous wavelet transform (CWT) to construct a time-frequency representation of structural responses and employs the DCRN to capture local non-uniform deformations and key feature regions adaptively. A Teacher-Student consistency-based semi-supervised learning mechanism is then applied to effectively exploit latent information from unlabeled data under limited labeled conditions. Additionally, the imbalance-aware loss function combining focal loss, logit adjustment, and distribution alignment is designed to suppress majority-class bias and enhance the stability of minority-class recognition. Finally, numerical simulations and ablation experiments on two datasets demonstrate that, even with only approximately 10% labeled samples and highly imbalanced class distributions, the proposed method achieves accurate recognition of mild, moderate, and severe damage states, with particularly superior performance in identifying severely damaged samples compared to traditional fully supervised models and semi-supervised methods.