For safety reasons, in order to ensure that a robot can make a reasonable response after a collision, it is often necessary to localize the collision. The traditional model-based collision localization methods, which are highly dependent on the designed observer, are often only useful for rough localization due to the bias between simulation and real-world application. In contrast, for fine collision localization of small-scale regions, data-driven methods can achieve better results. In order to obtain high localization accuracy, the data required by data-driven methods need to be as comprehensive as possible, and this will greatly increase the cost of data collection. To address this problem, this article is dedicated to developing a data-driven method for zero-shot collision localization based on local region data. In previous work, global region data were used to construct the collision localization model without considering the similarity of the data used for analysis caused by the assembly method of the contact parts. However, when using local region data to build collision localization models, the process is easily affected by similarity, resulting in a decrease in the accuracy of collision localization. To alleviate this situation, a two-stage scheme is implemented in our method to simultaneously isolate the similarity and realize collision localization. Compared with the classical methods, the proposed method achieves significantly improved collision localization accuracy.