With the rapid development of the Internet and widespread use of social media, massive amounts of multimodal data are generated daily on various platforms. Effectively and accurately detecting and identifying social events from such data has become a critical challenge in public opinion analysis and social governance. Traditional methods rely on fixed category sets and abundant labeled data, making them difficult to adapt to real-world scenarios. First, social media data exhibit pronounced temporal dynamics, with significant differences in distribution and semantics across different time periods. Second, new event types continuously emerge, meaning that data in the current time period include both known and unknown categories, and the total number of categories is often unknown in advance. Third, in practical applications, it is often difficult to obtain or revisit source-domain data from historical time periods, which requires the model to adapt to the target time period under source-free conditions. To address these challenges, this study proposes an open-world multimodal social event detection method based on source-free universal domain adaptation. The proposed method leverages a model trained on a previous time period as prior knowledge and performs subspace decomposition of the multimodal feature space on data from the current time period, based on prior classifier weights. The feature representations are separated into subspaces associated with known semantics and potential unknown semantics, within which synthetic samples are generated for training the target classifier. During the target time period, the feature extractor and classifier are jointly optimized. By integrating unbalanced optimal transport with classifier prediction distributions, a sample-level confidence consistency filtering mechanism is constructed. High-confidence samples are used for pseudo-label supervision to enhance the discriminative ability of known categories, while low-confidence samples are guided by structural consistency constraints to form stable structures in the feature space. As the feature structures are progressively optimized, both the discriminative ability of the classifier and the reliability of sample confidence estimation are further improved, thus forming a positive feedback loop in which discriminative ability, confidence filtering, and structural optimization mutually reinforce each other. Experimental results on two public datasets demonstrate that the proposed method outperforms the baseline methods, validating its effectiveness.