This paper presents a fault diagnosis approach based on Bayesian network for wide area backup protection of power system， which uses the activate signals of traditional measuring elements of relays as symptom information. In accordance with information quantity， size of diagnosis model and real-time request of diagnosis， a transmission network is divided into diagnosis zones. The stochastic characteristic of false operation and failure operation of relays is analyzed， and fault diagnosis models for each diagnosis zone are built. Test results for a sample power system show that the developed fault diagnosis model is correct and efficient， and has strong fault-tolerant ability.