<p>Wearable sensor–based human activity recognition (HAR) is central to smart sport and health-monitoring applications, yet existing deep learning models often rely on ad hoc multimodal fusion and offer limited interpretability. This paper proposes FusionProtoNet, an interpretable deep learning framework for multimodal wearable HAR that integrates three key components: Location-Graph Attention Fusion (LoGAF) for placement-aware cross-location fusion, a Temporal Conformer encoder for joint local and long-range temporal modeling, and a Prototype Reasoning Head for case-based interpretability. LoGAF is a placement-aware fusion module that explicitly learns how signals from different body locations (wrist, chest, ankle, and heart rate) should interact, rather than simply concatenating them. FusionProtoNet is evaluated on the PAMAP2 Physical Activity Monitoring dataset using a subject-wise cross-validation protocol. Experimental results demonstrate that the proposed model achieves 97.9% classification accuracy and a macro-averaged AUC of 99.4%, outperforming strong CNN, LSTM, and Transformer-based baselines. In addition to improved predictive performance, FusionProtoNet provides transparent explanations through temporal saliency visualization, channel-level attribution, and prototype retrieval, confirming that its decisions are grounded in biomechanically meaningful sensor cues. These results indicate that FusionProtoNet advances both accuracy and interpretability for wearable HAR, supporting trustworthy deployment in smart sport, rehabilitation, and health-monitoring scenarios.</p>