This study introduces a novel representation learning method to enhance unsupervised deep clustering in Human Activity Recognition (HAR). Traditional unsupervised deep clustering methods often struggle to extract effective feature representations from unlabeled data, failing to fully capture the true underlying structure of the data. As a result, classification performance is frequently suboptimal. To address this limitation, we propose leveraging an autoencoder integrated with models pre-trained on diverse HAR datasets to extract robust and transferable feature representations from target data. These representations are subsequently utilized within an unsupervised deep clustering framework, enabling effective discovery of the data’s latent structure and significantly improving clustering performance. The proposed method was evaluated on three HAR datasets and compared against conventional approaches, including autoencoder-based deep clustering and traditional classification methods such as k-means and HMM. As a result, the proposed method achieved F1 scores ranging from 0.441 to 0.781, significantly outperforming the baseline scores of 0.215 to 0.459. Furthermore, with fine-tuning using only 50 samples, the proposed method achieved even higher accuracy, with F1 scores ranging from 0.66 to 0.88. Additionally, it exhibited higher accuracy and robustness compared to traditional classification methods, highlighting its effectiveness in unsupervised learning scenarios. This study not only advances recognition accuracy in HAR but also demonstrates the potential of cross-dataset representation learning to effectively utilize unlabeled data. The proposed method offers a scalable and practical solution with broad applicability beyond HAR to other domains.