Nowadays, people's expression on the Internet is no longer limited to text, especially with the rise of the short video boom, leading to the emergence of a large number of modal data such as text, pictures, audio, and video. Compared to single mode data ,the multi-modal data always contains massive information. The mining process of multi-modal information can help computers to better understand human emotional characteristics. However, because the multi-modal data show obvious dynamic time series features, it is necessary to solve the dynamic correlation problem within a single mode and between different modes in the same application scene during the fusion process. To solve this problem, in this paper, a feature extraction framework of the three-dimensional dynamic expansion is established based on the common multi-modal data, for example video , sound ,text.Based on the framework, a multi-modal fusion-matched framework based on spatial and temporal feature enhancement, respectively to solve the dynamic correlation within and between modes, and then model the short and long term dynamic correlation information between different modes based on the proposed framework. Multiple group experiments performed on MOSI datasets show that the emotion recognition model constructed based on the framework proposed here in this paper can better utilize the more complex complementary information between different modal data. Compared with other multi-modal data fusion models, the spatial-temporal attention-based multimodal data fusion framework proposed in this paper significantly improves the emotion recognition rate and accuracy when applied to multi-modal emotion analysis, so it is more feasible and effective.