To address the problems of poor differentiation of soft fault features of power electronic circuits and not easy to diagnose, a fault diagnosis method of variational modal decomposition ( VMD) combined with an improved sparrow search algorithm ( ISSA) optimized extreme learning machine (ELM) is proposed. Firstly, the acquired fault signals are decomposed into the intrinsic modal components (IMF) by VMD, and the twelve-dimensional time-domain parameters of the linearly reconstructed IMF are extracted as the feature vectors for fault diagnosis. Secondly, in order to improve the accuracy of ELM in fault diagnosis, ISSA is proposed to optimize the parameters of ELM and establish ISSA-ELM classification model. ISSA is improved by three strategies such as initializing the population with Iterative mapping, introducing adaptive inertia weight factor at the discoverer position update, and introducing levy variation operator to perturb at the solution position to get a new solution to improve the algorithm performance. In the 8-class benchmark function test, ISSA has improved the convergence speed and finding accuracy than the other 4 intelligent algorithms, and the accuracy of VMD combined with ISSA-ELM reaches more than 99% in the soft fault diagnosis of 150 W Boost circuit.