Through-Silicon Vias (TSV) is a key technology enabling three-dimensional packaging, which has received much attention due to its unique vertical interconnect. However, the complexity of the silicon via process increases the chances of defects, which are not easy to be detected, thereby affecting the performance and reliability of the packaging. Consequently, a dynamic excitation-based internal detection method is proposed in this paper. By applying dynamic thermal excitation to the packaged chip, internal defects are stimulated to produce abnormal temperature distributions, and time series images of the temperature distribution on the outer surface of the package are collected, which are recognized and classified using deep learning to achieve external diagnosis of the internal defects. Firstly, a three-dimensional packaging model is constructed for transient thermal finite element simulation, and the simulation analysis reveals that internal defects have subtle impacts on the temperature distribution. Then, a three-dimensional Convolutional Neural Network (C3D) model is constructed to recognize and classify defects by analyzing the temporal changes in temperature gradient distribution images. Finally, an experimental testing platform is established, and three-dimensional packaging samples containing various defects are prepared for validation. The results show that the classification accuracy of the dynamic excitation-based internal defect detection method can reach up to 97.81 %, offering a new perspective for the detection of internal defects in three-dimensional packaging.