The invention discloses a community self-organizing detection method for power network fault diagnosis. The method comprises the steps of firstly, collecting network characteristic parameters of power networks, then describing the power networks as weighted network models, constructing local fitness and global fitness functions, starting from grouped solutions of the power networks, which are generated randomly, calculating local fitness of each power node, sequencing the local fitness, selecting the nodes with the poor local fitness according to an expansion evolution probability distribution function, transferring the nodes with the poor local fitness to another group of networks to generate new solutions, comparing global fitness values of the new solutions and the current solutions, reserving the best solutions in the new solutions and the current solutions, enabling the new solutions to serve as initial solutions for the next iteration to repeat above optimization processes until preset end conditions are met, and finally, analyzing and outputting community self-organizing detection results which are used for power network fault diagnosis. Compared with conventional methods, the method has the advantages of being a few in adjusting parameter, simple in detection process, easy to implement and high in detection efficiency and detection precision.