The combination of dissolved gas analysis and FCM clustering is effective in improving the accuracy rate of power transformer fault diagnosis, but the result of FCM clustering is unstable and easy getting stuck in a local optimum. Optimized FCM clustering by the proposed CSO (CSO-FCM) is introduced to diagnose the fault of transformer in order to conquer the shortages of FCM clustering. The CSO algorithm includes horizon cross as well as vertical cross, whose combining enhances the global convergent ability while the introduction of competitive mechanism drives the potential solutions approximate the global optima in an accelerating fashion without sacrificing the convergence speed. This method effectively compensates the demerits of single intelligent algorithm, which not only has the ability to dispose unstable information of fuzzy theory, also has an advantage of global convergence of CSO. Simulation and case analysis indicate that, compared with the traditional FCM clustering, the CSO-FCM clustering can obtain high performance clustering center and effectively raise power transformer fault diagnosis accuracy and diagnosis speed.