As a key indicator of teachers’ teaching ability in digital environment, digital literacy covers many aspects such as information acquisition, processing, analysis and innovative application. The traditional evaluation system often ignores this comprehensive index and adopts a single quantitative or qualitative model, which is difficult to fully reflect the true performance of teachers. In this study, multimodal data fusion technology is introduced to integrate multimodal data such as teaching videos, teaching logs, student feedback, etc., features are extracted and fused by deep learning algorithm, and weights are dynamically adjusted by reinforcement learning to realize real-time and comprehensive evaluation of teachers’ digital literacy. The experimental results show that the accuracy of the fused features is significantly improved, and the scores of teachers’ digital literacy are highly correlated with the actual performance, and the scores show an upward trend with time, reflecting the dynamic improvement of teachers’ digital literacy. In addition, the personalized feedback report has been highly recognized by teachers, effectively helping teachers improve their digital literacy. Compared with the traditional evaluation methods, this algorithm is excellent in accuracy, comprehensiveness and dynamics, which provides a new scientific basis and method for the evaluation of university teachers’ digital literacy, and also provides a reference for multimodal data fusion and evaluation in other fields.