This paper addresses the challenges in dance action recognition (DAR) by proposing an optimization method and an improvement strategy to enhance the accuracy and stability of DAR. First, we design a deep learning model combining a convolutional neural network (CNN) with a self-attention mechanism for identifying dance movements. Experimental results show that the new method significantly improves accuracy on the test set, outperforming traditional methods. We also perform a sensitivity analysis of key parameters, demonstrating that proper parameter tuning can further enhance model performance and stability. Additionally, the method performs well in recognizing various dance movements, achieving high accuracy and generalization ability. In conclusion, this study provides a theoretical foundation and experimental validation for optimizing dance action recognition and evaluation technology, offering valuable technical support for the intelligent development of the dance field.