As educational technology evolves, personalized and adaptive learning is becoming increasingly important. Knowledge Tracing (KT), a key technique in achieving this, updates students’ knowledge states in real time by analyzing their interactions with educational systems. Existing models usually focus on a single perspective (problem or concept), which may result in an inability to fully capture the dynamic changes in students’ knowledge state. And when dealing with complex knowledge structures, how to continuously and dynamically update students’ knowledge mastery while improving the accuracy of prediction is also a challenge. In order to solve these problems, we proposed a novel knowledge tracing method based on multi-perspective dynamic evolution (DMKT). The model tracks students’ knowledge state from three levels: problem, concept, and overall, to comprehensively capture the progression of their knowledge. DMKT uses LSTM to capture changes in knowledge state, with the IRT model predicting student performance and enhancing model interpretability. The self-attention mechanism updates knowledge state, capturing long-term dependencies. Experimental results demonstrate that DMKT outperforms the latest baselines in terms of prediction accuracy and enhances the performance of knowledge state tracking and updating.