Combination of active sensing method and data-driven approach for rubber aging detection

Abstract

Rubber bearings are key components of base-isolated structures, and the monitoring of their damage states is an important task. Aging is a primary concern affecting the service life and isolation effect of rubber bearings. Therefore, this study combined an active sensing method and a data-driven approach to detect rubber aging. A shear stiffness, accelerated aging, and active sensing experiments were conducted on a scaled rubber specimen. As the aging level increased, the shear stiffness of the specimens gradually increased from 116.69 to 127.82 N/mm, but this change was not linear. Due to variations in the degree of aging, discrepancies may arise in the time and frequency domain characteristics of detection signals. However, establishing an empirical relationship between the degree of aging and the features of detection signals were highly challenging. A deep-learning-based data-driven method was used to predict the aging level and shear stiffness using detection signals. The deep learning model successfully detected the aging level, and the prediction accuracy on the validation and test sets reached 99.98%. For the deep learning model for aging level prediction, the optimal input vector length is 4096, the recommended number of layers is 3–5, and the recommended number of cells in each layer is 256–2048. Moreover, the deep learning model also detected the shear stiffness of the rubber specimen. The mean absolute error was 0.27 N/mm on the validation set and 0.28 N/mm on the test set. For the deep learning model for shear stiffness prediction, the optimal input vector length is 4096, and the optimal structure is seven layers with 2048 cells in each layer.