Fault diagnosis in power grids is known to be challenging, due to the massive scale and spatial coupling therein. In this study, we explore multiscale network inference for fault detection and localization. Specifically, we model the phasor angles across the buses as a Markov random field (MRF), where the conditional correlation coefficients of the MRF are quantified in terms of the physical parameters of power systems. Based on the MRF model, we then study decentralized network inference for fault diagnosis, through change detection and localization in the conditional correlation matrix of the MRF. Particularly, based on the hierarchical topology of practical power systems, we devise a multiscale network inference algorithm that carries out fault detection and localization in a decentralized manner. Simulation results are used to demonstrate the effectiveness of the proposed approach.