Abstract
Sports behavior prediction requires precise and reliable analysis of muscle activity during exercise. This study proposes a multi-channel correlation feature extraction method for electromyographic (EMG) signals to overcome challenges in sports behavior prediction. A wavelet threshold denoising algorithm is enhanced with nonlinear function transitions and control coefficients to improve signal quality, achieving effective noise reduction and a higher signal-to-noise ratio. Furthermore, multi-channel linear and nonlinear correlation features are combined, leveraging mutual information estimation via copula entropy for feature construction. A stacking ensemble learning model, incorporating extreme gradient boosting (XGBoost), K-nearest network (KNN), Random Forest (RF), and naive Bayes (NB) as base learners, further enhances classification accuracy. Experimental results demonstrate that the proposed approach achieves over 95% prediction accuracy, significantly outperforming traditional methods. The robustness of multi-channel correlation features is validated across diverse datasets, proving their effectiveness in mitigating channel crosstalk and noise interference. This work provides a scientific basis for improving sports training strategies and reducing injury risks.