 This paper proposes a novel approach to daily activity recognition in smart homes. The authors first unified the sensor space and activity space to reduce variability across heterogeneous environments. Next, they used the Word2Vec algorithm to convert activity samples into digital vectors recognizable by the network. Finally, the deep network was fine-tuned to transfer knowledge and complete the recognition task. This approach improved the accuracy of the recognition task while reducing the time cost and need for heavy data annotation. Furthermore, the authors demonstrated that their approach could be applied to public datasets with limited data, achieving comparable performance to networks trained on larger datasets. This article was authored by Yu Jinyu, Kuen Tang, and Yaxing Liu.