<p>Human activity recognition (HAR) has been identified as playing a pivotal role in clinical and health monitoring applications. However, extant methodologies frequently encounter threefold challenges: existing models ignore temporal heterogeneity, treating multi-channel data as homogeneous and causing feature entanglement. Heuristic multi-scale stacking fails to explicitly decompose trends, periods. Fragmented learning across time, frequency, and decomposition domains forces unified models to fit conflicting patterns-explaining why deeper Transformers underperform, as self-attention cannot adapt to distinct temporal dynamics. In order to address these issues, the present paper puts forward a novel temporal decomposition attention network (TDA). The model under consideration makes a departure from homogeneous signal processing constraints by adopting a “decomposition-specialization-fusion” paradigm. This enables the explicit modeling of the intrinsic structure of sensor time-series data. The multi-scale decomposition module is specifically designed to separate the input signal into trend and cyclical components. The trend decomposition attention module utilizes lightweight global dependency modeling to identify long-term trend components. The cyclical decomposition attention module enhances cyclical patterns through frequency-domain augmentation and spatiotemporal collaborative attention mechanisms. Furthermore, an adaptive gated fusion mechanism integrates multi-domain features to generate dynamic representations with high discriminative power and interpretability. Extensive experimentation on public benchmark datasets such as UCI-HAR, WISDM, and PAMAP2 has demonstrated the outstanding performance of TDA in activity recognition tasks. The model demonstrates notable robustness and generalization capabilities across users and devices. This research not only advances activity recognition performance, but also pioneers a novel paradigm based on temporal data analysis.</p>