Abstract: Concrete-filled steel tube (CFST) columns have been widely used in the canopy structures of large-span, heavy-load railway stations due to their outstanding load-bearing capacity. However, owing to factors such as concrete shrinkage, creep, and construction techniques, voids of varying degrees may develop between the steel tube wall and the internal concrete. This delamination (or “hollowing”) phenomenon significantly reduces the composite load-bearing performance of CFST members, thereby posing a serious threat to structural safety. To address this issue, this paper proposes an intelligent recognition method for internal void defects in CFST members based on the Particle Swarm Optimization–Support Vector Machine (PSO-SVM) algorithm. Ultrasonic testing was conducted on both intact and hollow CFST specimens. The acquired ultrasonic signals were processed using wavelet transform, and key time–frequency feature parameters were extracted to construct a labeled dataset. The SVM classifier was then optimized using the PSO algorithm. The optimized classification model achieved a recognition accuracy of 98.46%. A comparative analysis was performed between the proposed PSO-SVM model and conventional classification models (e.g., standard SVM, BP neural network). Results demonstrate that the PSO-SVM model outperforms these alternatives in the intelligent classification and recognition of internal void defects in CFST members, enabling effective, automated defect identification and providing robust technical support for engineering inspection and maintenance.