Deep-SVDD-based Real-time Early Warning for Cable Structure

Abstract

Cables are essential load-bearing components in many structures, and early warning methods are crucial for ensuring their safety. Early warning methods based on structural dynamic responses have played a key role due to their low cost, easy maintenance, and replaceability. However, extracting robust damage-sensitive features from dynamic responses under ambient excitation remains challenging. In this paper, an early warning method based on deep support vector data description is proposed. The method extracts damage features from the power spectral density of lateral acceleration of cables and interpretable damage indicators are proposed for identifying cable interaction. The proposed unsupervised learning method only requires healthy state lateral acceleration data for model training. Numerical and field experiments on the Shanghai-Suzhou-Nantong Yangtze River Bridge demonstrate the method's effectiveness in early warning for cables. Compared to the deep auto-encoder based damage diagnosis method, the proposed method shows higher accuracy and potential for real-time early warning of cable structures.