We propose a fault diagnosis method based on the modified cuckoo search algorithm (WCS), steepest descent method and support vector machine (SVM) to improve the accuracy of transformer fault diagnosis. A new inertia weight is also proposed and applied to solve the problem that the convergence rate of cuckoo search algorithm decreases in final iterations. SVM parameters are optimized by the algorithm which is combined with improved cuckoo search algorithm and steepest descent method, overcoming the defects that SVM model is easy to fall into local optimum. Support vector machine is trained on the MATLAB platform using LIBSVM toolbox, and the well-trained SVM will be adopted to diagnose the #1 transformer fault for 110kV Gantang substation. Study of practical cases indicate that, with this method, transformer faults can be diagnosed effectively and accurately, and the accuracy is higher than that using particle swarm optimization(PSO)、genetic algorithm(GA) and grid search(GS).