The integrity of Concrete-Filled Steel Tubular (CFST) columns can be significantly compromised by subsurface voids or defects, which pose substantial safety risks. This study proposes a novel Dual-Branch Convolutional Neural Network (CNN) Framework with Multi-modal Feature Fusion for detecting subsurface defects in CFST columns. The proposed method leverages the strengths of CNNs in feature extraction and combines multiple modalities of data, such as time-domain and Gramian Angular Field (GAF) image features, to enhance detection performance. By employing a dual-branch structure, the model can capture both temporal and spectral characteristics of percussion signals, leading to more accurate and robust predictions. Field experiments on high-rise building CFST columns demonstrate that the percussion method, validated by thermal imaging, effectively identifies void locations with an accuracy of 94.5%. The results indicate that the proposed method outperforms traditional techniques, providing a reliable solution for structural health monitoring and ensuring the safety and longevity of CFST columns.