Research on the construction of basketball training behavior analysis system based on machine learning——A cross-national empirical study on the digital transformation of physical education in colleges and universities

We introduce FL-BTBAS, a federated learning framework designed for analyzing basketball training behaviors, tackling non-IID data distribution and privacy concerns in decentralized sports data environments. By combining adaptive dynamic pruning with a multimodal CNN-LSTM framework, it facilitates cross-facility collaborative training without sharing raw data: local clients handle RGB-D and inertial sensor data using layer-wise magnitude pruning, which adjusts sparsity to address data heterogeneity, while the global model integrates pruned structures through weighted averaging and structural alignment to manage variations among clients. Combining pose estimation with biomechanical data at the edge cuts latency by 40% compared to cloud-based approaches. FL-BTBAS, deployed on NVIDIA Jetson AGX Orin, incorporates heterogeneity-aware sparsity management and federated structural alignment to maintain the significance of underrepresented patterns, showcasing scalability, privacy adherence, and precise analytics for multi-institutional digital physical education.