<p>In multi-sensor behavior recognition, data redundancy and irrelevant features severely impair model training and optimization, leading to degraded performance. Feature selection enhances model performance by reducing dimensionality and retaining representative features. Existing dynamic selection methods iteratively update feature weights via forward propagation but are sensitive to initial parameters, which may result in early selections that reflect only local feature importance. This issue hinders the characterization of global contributions, undermines subset representativeness, and impairs model generalization. To address these challenges, we propose a Two-Stage Dynamic Feature Selection Approach (TSDFSA) based on attention mechanisms. TSDFSA adopts a two-stage feature selection strategy comprising a pre-training stage and a dynamic selection stage. The pre-training stage alleviates the impact of initial parameter settings on the feature selection process. In the dynamic selection stage, a soft–hard coupling gating mechanism is employed to adaptively adjust feature weights. This mechanism progressively refines the feature subset and avoids the limitations of one-off selection. The Hadamard product-based weighting strategy further improves the discrimination of feature contributions. Subsequently, a binary feature mask is constructed using one-hot encoding to eliminate irrelevant features. Experimental results on two multi-sensor datasets demonstrate that TSDFSA significantly enhances behavior recognition accuracy, achieving 94.10 and 91.01, respectively. Compared to other deep learning-based feature selection methods, TSDFSA achieves superior performance, confirming its effectiveness in both feature selection and behavior recognition.</p>