Abstract: In order to diagnose several common types of switch faults and reduce the loss caused by misclassification, this paper presented a new switch fault diagnosis system which was based on the genetic algorithm based cost sensitive RBF neural network model. In this model, the cost sensitive fitness function was established to guide the optimal direction of random search based on genetic algorithm, which has lower false classification cost. The system was validated by using the monitoring data of switch action current at a certain railway station. The result showed that this system could not only improve the overall recognition accuracy of the fault data, but also significantly reduce the proportion of false classification of fault data. So this system could help railway maintain personnel quickly and accurately diagnosis the switch fault, reduce the fault processing time, and improve the safety of railway traffic.