In view of the limitations of current digital learning platforms in personalized recommendation and dynamic knowledge state modeling, this study proposes an autonomous learning platform architecture based on hierarchical dynamic Bayesian network (HDBN), which achieves high-precision adaptive learning by integrating knowledge graph embedding and multimodal behavior data analysis. The platform adopts a hierarchical architecture design. The data layer builds a multi-source behavior data collection system based on the xAPI specification, the modeling layer uses HDBN for knowledge state modeling, and the application layer includes a personalized recommendation engine and a learning path optimization module. The user behavior data is reduced in dimension by using a variational autoencoder, and the approximate reasoning of large-scale networks is realized by combining the Loopy Belief Propagation algorithm. The average test pass rate of the experimental group (based on the HDBN self-learning platform) reached 90.52%, the attendance rate is excellent, and the average skill achievement rate is 92.544%. The platform constructed in this paper provides new ideas and methods for the intelligent development of digital learning platforms.