Signal Angle-of-Arrival estimation has been widely studied as an important sensing technology in Integrated Sensing and Communications, but its performance is limited by complex multipath environments and the scale of antenna arrays. In recent years, learning-based signal angle-of-arrival estimation methods have made breakthroughs, but the effectiveness of learning methods is closely related to the richness of datasets. In the new generation of communication systems, the large-scale deployment of base stations for sensing needs, the data set is difficult to cover a large number of complex and changeable scenarios, and there are problems such as excessive collection scale and difficult labeling. In order to achieve effective estimation of signal angle-of-arrival in complex environments, this paper uses the method of transfer learning to reduce the cost of data set collection and improve the accuracy of estimation. The transfer learning in this paper uses a domain adversarial model to achieve unlabeled transfer from simple to complex scenes. The experiments set up multiple scenarios such as multi-source, multi-path, and array defects. The migration results show that the method can effectively improve the performance of angle-of-arrival estimation in most scenarios, and verify the adaptability of the proposed method to complex environments.