Purpose This study aims to address the limitations of supervised learning in structural damage detection by proposing an unsupervised method that eliminates the need for large, well-labeled data sets. The method uses global transmissibility within an autoencoder framework to detect structural damage effectively. Design/methodology/approach A novel pooling-free Deep Convolutional Autoencoder architecture with batch normalization and ReLU activation is developed to extract damage-sensitive features from global transmissibility functions. Additionally, a denoising technique as CEEMDAN-CNSPCA is integrated to enhance performance under noisy conditions. Validation is conducted using numerical simulations and experiments on beam, steel tower and steel frame models. Findings The proposed method demonstrates high damage detection accuracy and robustness across various structures and damage scenarios. Global transmissibility is shown to capture structural behavior more comprehensively than local transmissibility. The integrated denoising technique significantly improves detection performance in high-noise environments. Research limitations/implications While the method shows promise, it requires further testing on diverse structural types and real-world operational conditions to generalize its applicability. The dependency on the quality of transmissibility data also warrants consideration. Practical implications The approach provides a viable solution for real-time structural health monitoring without the need for extensive labeled data sets, reducing both cost and time. Its robustness to noise makes it suitable for deployment in field environments. Social implications Reliable, unsupervised damage detection techniques can enhance public safety by facilitating early detection of structural failures in critical infrastructure without intensive manual inspection. Originality/value This study introduces a novel damage detection method using global transmissibility within an autoencoder framework. This unsupervised learning approach eliminates the reliance on large, well-labeled data sets and demonstrates good performance across multiple structural types, including beams, steel towers and frames.
Qiao et al. (Tue,) studied this question.