Rapid sealant damage assessment in hidden frame glass curtain walls via machine learning and dynamic testing

Abstract

Hidden frame glass curtain walls are widely used in modern buildings. The damage of structural sealants in curtain walls under long-term service may pose potential safety concerns. Motivated by the need for rapid damage assessments of structural sealants in glass panels, this study proposes a data-driven approach that integrates on-site dynamic testing with a machine learning methodology. The machine-learning model is trained via a conservative strategy to reduce underestimated damage and via a simulated dataset comprising 75,000 finite-element models of glass panels with various sealant damage. Comparisons between multiple machine-learning models reveal that eXtreme Gradient Boosting achieves the best performance, with an accuracy of 85.96% in classifying the conditions of glass panels. An interpretability analysis shows that the relative change of the first-order natural frequency is the most critical feature governing the classification task. A dynamic testing method incorporating both multi-point and sparse sensor allocations is proposed. The multi-point accelerometers allocation identifies the modal properties of glass panels. To accelerate assessment, a sparse allocation, which employs a laser vibrometer, is utilized to determine the modal frequencies of glass panels with identical specifications. The proposed approach is validated via field tests in high-rise buildings. The sealant damage assessment of actual glass panels by the approach agrees well with manual inspection results.