With the development of artificial intelligence, machine learning algorithms are currently being used in more and more fields, such as autonomous driving, medical diagnosis, etc. In recent years, much research focuses on property verification of machine learning models. As one of the machine learning models, the tree ensemble model's structure is amicable to formal verification, but large models still prove hard to verify due to the combinatorial path explosion. This paper presents a violation-driven, sound but incomplete method from a testing perspective. We generate an explanation model of the original model and verify it formally. After a narrowed search space is obtained, we verify the original model by a testing-based method. A counterexample is then proof that the original model violates the property. We elaborate our method through a case study in detail. And we have developed our method into a tool called TEPV (Tree Ensemble Property Verification) and tested it on datasets of various sizes. The experiment demonstrates that our approach is scalable and works well on large tree ensemble models.