Dong Kaiyuan, Abdullah Borhannudin Bin, Bin Abu Saad Hazizi, Lu Chenxi
Department of Sports Studies, Faculty of Educational Studies, Universiti Putra Malaysia, Serdang, 43400, Selangor, Malaysia.
Department of Nutrition, Faculty of Medicine and Health Sciences, Universiti Putra Malaysia, Serdang, 43400, Selangor, Malaysia.
Sci Rep. 2025 Aug 17;15(1):30124. doi: 10.1038/s41598-025-16240-w.
This study aims to improve both the evaluation accuracy and the real-time feedback capability in monitoring athletes' physical function changes during volleyball training. Firstly, based on the framework of the generalized regression neural network, a variable-structure generalized regression neural network (VSGRNN) is proposed and developed. Three heterogeneous kernel functions, namely Gaussian kernel, radial basis kernel, and Matern kernel, are introduced, and a local weighted response mechanism is constructed to enhance the expression ability of nonlinear physiological signals. Second, a dynamic adjustment mechanism for smoothing factors based on local gradient perturbation is proposed, enabling the model to have response compression capability in high-fluctuation samples. Finally, combining the structure embedding mapping mechanism with a multi-scale linear compression framework, the reconstruction of high-dimensional physiological indicators and the elimination of redundant features are achieved, improving model deployment efficiency. Comparative experiments conducted on training data of a high-level university men's volleyball team show that VSGRNN has a goodness-of-fit R = 0.927 on the validation set, with a Root Mean Square Error (RMSE) only 1.68 and Symmetric Mean Absolute Percentage Error (SMAPE) controlled at 8.21%. Within the local perturbation interval, the peak response deviation is 6.7%, far better than the comparative models (Long Short-Term Memory (LSTM) + Attention at 8.5% and Tabular Data Network (TabNet) at 9.8%). When compressed to 30% of the original feature dimension, the error only increases by 7.9%, and the inference time is shortened by 46.1%. The research conclusion shows that VSGRNN outperforms traditional models in terms of accuracy, robustness, structural compression adaptability, and real-time feedback capability. This study provides an engineerable structure-response modeling method for the intelligent evaluation of physical functions in volleyball-specific training, which has high practical application value.
本研究旨在提高排球训练期间监测运动员身体机能变化的评估准确性和实时反馈能力。首先，基于广义回归神经网络框架，提出并开发了一种变结构广义回归神经网络（VSGRNN）。引入了高斯核、径向基核和马特恩核三种异构核函数，并构建了局部加权响应机制以增强非线性生理信号的表达能力。其次，提出了一种基于局部梯度扰动的平滑因子动态调整机制，使模型在高波动样本中具有响应压缩能力。最后，将结构嵌入映射机制与多尺度线性压缩框架相结合，实现了高维生理指标的重构和冗余特征的消除，提高了模型部署效率。对某高水平大学男子排球队的训练数据进行的对比实验表明，VSGRNN在验证集上的拟合优度R = 0.927，均方根误差（RMSE）仅为1.68，对称平均绝对百分比误差（SMAPE）控制在8.21%。在局部扰动区间内，峰值响应偏差为6.7%，远优于对比模型（长短期记忆网络（LSTM）+注意力机制为8.5%，表格数据网络（TabNet）为9.8%）。当压缩到原始特征维度的30%时，误差仅增加7.9%，推理时间缩短46.1%。研究结论表明，VSGRNN在准确性、鲁棒性、结构压缩适应性和实时反馈能力方面优于传统模型。本研究为排球专项训练中身体机能的智能评估提供了一种可工程化的结构-响应建模方法，具有较高的实际应用价值。
