Dynamic binary translation plays a crucial role in cross-platform portability and compatibility. However, existing methods face challenges in translation efficiency and resource overhead. A machine learning-based dynamic binary instruction set translation method is proposed. The proposed method innovatively integrates machine learning-based binary analysis and neural machine translation into a dynamic binary translation system, constructing an efficient and accurate instruction set translation framework. The method includes leveraging pre-trained models to capture instruction-level semantic information, training a Transformer-based neural translation model for instruction mapping, and integrating machine learning-based mappings with dynamic binary translation systems through formal verification. Experimental results from SPEC CPU 2006 benchmark tests demonstrate that the proposed method achieves superior translation rule coverage and runtime efficiency compared to existing approaches, offering new insights for optimizing dynamic binary translation systems.