With the rapid development of vocational education, there is an increasing demand for intelligent systems that can provide real-time feedback and personalized guidance in practical skills training. Traditional training methods relying on repetitive practice and instructor supervision are inefficient and lack precision. To address these challenges, this paper proposes DualStreamNet (DSNet), an innovative action recognition method designed to enhance the quality of practical teaching in higher vocational education. DSNet integrates spatial and temporal convolutions within a unified framework, enabling simultaneous modeling of skeleton data’s spatial topology and temporal dynamics. Additionally, an uncertainty estimation module is introduced to evaluate prediction credibility, improving the model’s robustness in handling ambiguous actions. Experimental results demonstrate that DSNet achieves superior accuracy and reliability compared to existing methods, providing a solid foundation for developing intelligent feedback systems in vocational education.