Multidimensional signal-aided machine learning for void detection in concrete-filled steel tubes✰

Concrete-Filled Steel Tube (CFST) arch bridges are highly vulnerable to debonding defects, which pose serious threats to structural safety. Existing detection methods are often limited by low accuracy, noise susceptibility, and a strong reliance on empirical judgment. To address these issues, this study proposes an intelligent debonding detection framework based on adaptive signal analysis and Machine Learning. The Grasshopper Optimization Algorithm (GOA) was employed to automatically optimize the parameters of Variational Mode Decomposition (VMD), establishing a GOA-VMD adaptive signal decomposition framework that effectively overcomes the challenges of mode mixing and parameter dependence inherent in traditional methods. Weak debonding features were further enhanced using the Teager Energy Operator, while high-resolution time–frequency analysis was achieved through the synchrosqueezed wavelet transform. A composite evaluation index combining Cyclic Kurtosis Entropy (CKE) was developed to construct a robust feature vector, which serves as the input for a Gradient Boosting Tree (GBT) classifier. Comparative analysis demonstrates that this feature-driven machine learning approach significantly outperforms conventional machine learning (Artificial Neural Network) models in terms of accuracy and robustness. Numerical simulations and feature analyses yielded the following findings: (1) a co-located measurement scheme (impact and sensor at the same point) improved detection stability by 72%; (2) measurement location was identified as the most influential factor for detection accuracy, contributing 38%; and (3) sensitivity for detecting debonding thicker than 4 mm reached 98%. The study establishes a reliable debonding identification technique that can be directly integrated into bridge health monitoring systems, offering an efficient tool for engineering applications.