Most link prediction models are black-box models with poor interpretability. Many scholars have proposed interpretability methods for link prediction. However, these methods often have limitations such as being tailored to a single target model, lacking generalizability, and having insufficient accuracy in the interpretation results. To address these shortcomings, this paper proposed a link prediction interpretation method based on edge perturbation. Firstly, the method used breadth-first search to obtain all paths from the head entity to the tail entity and then search for the neighboring nodes of the entities in these paths to form a training subgraph for the target triplet. Then, the method applied edge perturbation to the subgraph and retrained the embedding model to calculate the influence of each edge on the prediction result. Finally, bidirectional beam search identified the path with the greatest influence on the prediction result, which served as the explanation path for the target triplet. Experiments show that the proposed method outperforms most link prediction interpretation methods on public datasets, improving accuracy(ACC) by 2.3% and area under the precision-recall curve(AUPR) by 1.9% compared to the state-of-the-art methods. Additionally, interpretation experiments on biomedical datasets for the task of drug repurposing using link prediction techniques demonstrate good understandability and inspiration. The main contribution of this paper is the proposal of an effective interpretation method that does not depend on a specific model, which obtains the interpretation paths through edge perturbation and path search, making the interpretation of the results more intuitive and easier to understand, while providing support for knowledge graph applications in different domains.