This study designs and implements an intelligent motion posture system that integrates digital twin technology with deep learning for action analysis and technique optimization in badminton training. The system first employs OpenPose to extract skeletal keypoints, followed by confidence masking, centralization, normalization, and missing-data completion to enhance data quality and stability. On the action recognition side, temporal models including TCN, ST-GCN, and LSTM are evaluated to classify five core badminton movements. On the generative side, a conditional Transformer is adopted to produce corresponding standard skeletal sequences, which serve as reference templates for comparison with actual performance and for assessing technical deviations. Preliminary experiments on a self-constructed dataset show that the proposed system achieves strong performance in both classification accuracy and generative consistency, and provides real-time visualization and training feedback through a digital twin interface. Overall, this study verifies the feasibility of applying AI and digital twin technology to sports science applications, and demonstrates its potential for further extension and real-world deployment.