Abstract: Abstract: Unsupervised thresholding methods for structural damage identification in civil engineering require no labelled data, yet suffer from inaccuracies near threshold values. To address the false positives and false negatives inherent in unsupervised structural damage thresholding approaches near threshold thresholds, this study proposes a probabilistic structural damage identification approach based on a hypersphere ring description, termed VAEKL-RDDP. The method is developed upon the Deep Support Vector Data Description (Deep-SVDD) framework. It integrates a hypersphere ring description with a variational autoencoder (VAE) and Kullback–Leibler (KL) divergence, building upon the Deep Support Vector Data Description (Deep-SVDD) framework. First, the VAE is pre-trained to reconstruct structural acceleration responses. Then, KL divergence is incorporated to jointly optimize the pre-trained VAE encoder and the hypersphere ring model, enabling the extraction of reliable classification boundaries from the posterior distributions of latent features. Finally, structural damage is identified based on the constructed hypersphere ring, and the enclosed data are evaluated using a cumulative probability density approach. In tests involving progressive damage of the real Z24 bridge structure and full-scale vibration table experiments on a wooden pavilion, the proposed method achieved average improvements of 24.9% in accuracy and 36.7% in recall compared with the baseline autoencoder reconstruction (AE) method. Relative to benchmark approaches such as Deep-SVDD and the diffusion model–based interpolation prediction (IMDIFFUSION) method, it further achieved average gains of 19.5% and 31.2% in accuracy and recall, respectively. These results verify the effectiveness of the proposed framework in improving the precision of structural damage detection and reducing missed detections.

Key words: unsupervised structural damage identification, time series reconstruction, probabilistic damage evaluation, cumulative probability density, variational autoencoder

摘要： 摘 要: 土木工程结构损伤识别中无监督的阈值方法无需标注数据，但数据的不确定性导致阈值附近存在识别不准的问题。针对无监督结构损伤阈值方法在阈值附近存在误报和漏报的问题， 在深度支持向量描述(Deep-SVDD)方法的基础上提出基于超球环描述的概率性结构损伤识别(VAEKL-RDDP)方法。该方法以变分自编码器(VAE)为框架，利用Kullback-Leibler (KL)散度约束构造超球环。首先，预训练VAE，重建结构加速度响应；其次，引入KL散度，联合训练预训练的VAE编码器与超球环描述方法，在加速度数据特征后验分布中提取可靠分类边界；最后，依据分类边界构造超球环，依托超球环对结构进行损伤识别，对超球环内数据采用累积概率密度方法进行评估。在真实桥梁结构Z24桥的渐进性损伤和足尺木亭振动台试验测试场景中，相比自编码器重建(AE)的基线方法，所提出方法的准确率和召回率分别平均提高了24.9%和36.7%，相比Deep-SVDD和扩散模型插补预测(IMDIFFUSION)方法等对比方法，准确率和召回率分别平均提升了19.5%和31.2%，提高了损伤检测的准确性，降低了漏报可能性。

关键词: 无监督结构损伤识别, 时间序列重构, 概率性损伤评估, 累积概率密度, 变分自编码器