Audio-visual embodied navigation is a key research direction in embodied intelligence. This task requires agents to leverage visual and audio cues in 3D environments to perceive and locate target sound sources. Although multimodal navigation frameworks based on deep reinforcement learning (DRL) have demonstrated promising results, existing approaches fail to dynamically model the importance of feature channels during visual-audio feature fusion. We introduce a feature-channel attention module with adaptive weighting to amplify task-critical features. Experiments confirm superior navigation success rates and path efficiency over baselines.