Abstract: Power communication network is an indispensable and important part of power system， it’s the basis of power grid dispatching automation and production management modernization， and an important technical means to ensure the safe， economic and stable operation of power grid. The diagnosis of communication faults still depends on manual experience， which is difficult to meet the safety production needs of increasingly large and complex communication network. Methods based on rule engine or neural network gradually encounter bottlenecks in the application of production environment. It is difficult to train due to less samples， or work independently in production environment as a black box. To solve the above problems， this paper proposes an alarm clustering algorithm based on improved Markov-clustering and a fault diagnosis algorithm based on sequence similarity and OPTICS clustering， adapt to the current small sample scenario of fault data. On the basis of the above algorithm results， the fault diagnosis knowledge base and its iterative learning mechanism are constructed by using a small number of labels. It is verified by the data accumulated in actual production. The results show that the relevant algorithms and knowledge base have a good effect in dealing with actual production problems.

Key words: power communication, fault diagnosis, knowledge base, semi-supervised clustering