Transformer is an important electric equipment in power system, its running state plays an important role in the safe operation of the system, this article will study the price Sensitive mechanism and introduce it to Relevance Vector Machine, and put forward the price Sensitive Relevance Vector Machine (Cost - Sensitive learning Relevance Vector Machine, CS - RVM). The algorithm take the misdiagnosis loss minimum cost as the goal, according to the bayesian fault categories of risk theory predicts new samples. With typical examples verified the CS - RVM has high diagnostic accuracy, and can avoid fault diagnosis to some extent, high-risk failure misdiagnosed as low risk failure. On this basis, try to applied to transformer fault diagnosis, based on the CS - RVM oil-immersed power transformer fault diagnosis method, in order to overcome the existing transformer fault diagnosis method and did not take into account the issue of misdiagnosis cost difference, and the transformer fault diagnosis based on DGA data instance to the effectiveness of the proposed diagnostic method are verified.