Extensive research has been carried on Hardware Trojan (HT) detection. Most detection methods at the register transfer level (RTL) are based on supervised learning models. However, supervised methods require the training data to contain Trojan circuits and often involve additional efforts to address class imbalance, which limits their applicability in real-world scenarios. In this paper, we propose a HT detection method based on an unsupervised learning model that is trained using only normal signals. Experimental results show that the proposed model achieves an average accuracy of 97.9, demonstrating the feasibility of applying unsupervised learning to RTL hardware Trojan detection and offering a new perspective for future research in this field.