Acoustic source localization (ASL) technique is an important step for structural health monitoring (SHM). ASL in three dimensional (3D) structures is more challenging. The 3D acoustic source localization technology not only has important significance for the non-destructive monitoring of large-scale 3D structures, but also is indispensable for the spatial sound source localization problems. More unknown parameters, large number of sensors, limited known properties of the 3D structure and complex nonlinear equations greatly hinder the application and development of 3D acoustic source localization. Besides, when the acoustic signal propagates through different media in a 3D heterogeneous structure, it is refracted at the interfaces following Snell's law which makes the ASL in heterogeneous structures even more difficult. In this paper, the basic theoretical research on the acoustic source localization in 3D heterogeneous structures is carried out considering the refraction effect. A localization technique is first proposed based on the triangular pyramid sensor cluster using only the time difference of arrival (TDOA) and the location information of clusters. This technique can predict the velocity and acoustic source location with a relatively small number of sensors. A 3D finite element based numerical model of a heterogeneous structure made of two materials was analyzed to verify the proposed acoustic source localization technique. The results show the reliability of the proposed technique.