A method was proposed to address challenges in detecting debonding in concrete-filled steel tube (CFST) columns by analyzing acoustic signals. In this approach, wavelet time-frequency diagrams are initially derived through the application of continuous wavelet transform to acoustic signals acquired from tapping on columns. The wavelet time-frequency diagrams serve as data samples for training the MobileNetv2 convolutional neural network model in order to identify concrete debonding. The laboratory test results show that the MobileNetv2 model achieved 100 % accuracy for both the training and verification sets, and 99 % accuracy for the test set. Four CFST columns in a high-rise building under construction were also measured. The findings indicate a recognition accuracy ranging from 86.7 % to 90 % and an error rate that falls within an acceptable range. These results confirm the viability of the method outlined in this paper for detecting debonding in CFST columns using acoustic signals.