College students face problems of insufficient monitoring accuracy and a lack of personalized training intervention in daily exercise, which restricts the effective improvement of their physical health. To address this challenge, a Self-Attention and Entropy optimized Semi-Supervised Generative Adversarial Network (SAE-GAN) model is constructed to achieve high-precision recognition of aerobic exercise and optimization of training feedback. The SAE-GAN model strengthens local and global dependencies by introducing a feature Self-Attention (SA) module, and reduces category uncertainty by combining information entropy loss, thereby maintaining robust performance with limited labeled samples. Performance test results show that the SAE-GAN model outperforms existing semi-supervised methods under different labeled sample conditions. When there are only 100 labeled samples per category, the recognition accuracy of SAE-GAN reaches 71.7%. Intervention experiments further reveal that personalized training assisted by the model can significantly improve college students' speed, core strength, and lower limb explosive power. The research results prove that SAE-GAN can maintain high recognition performance under small sample conditions. Meanwhile, SAE-GAN can provide a feasible technical means for precise health intervention and expand the application prospects of artificial intelligence in sports science and college students' physical fitness improvement.
Su et al. (Thu,) studied this question.