With the continuous development of exoskeleton technology. HAR (human activity recognition) has been receiving increasing attention. It has been widely applied in fields such as healthcare, industry, and artificial intelligence. Currently, various machine learning algorithms have been utilized to achieve HAR using sensor data, resulting in promising outcomes. However, traditional machine learning methods are significantly influenced by the feature extraction process. To address these challenges, an INC-GRU model integrated with an attention mechanism was proposed. The model employs an Inception network and CBAM(convolutional block attention module) for spatial feature extraction. GRU (gated recurrent units) and an attention mechanism were used for temporal feature extraction. This approach effectively leverages both spatial and temporal information in time series data, allowing different weights to be assigned to each data point, thereby enabling selective focus on key information. Experiments conducted on the UCI-HAR and WISDM datasets have achieved F₁ scores of 96.64% and 97.92%, respectively. Comparative analyses with existing models have demonstrated the superiority of the proposed INC-GRU model.