The present study proposes a methodology for the automatic diagnosis of short-circuit faults in distribution systems using modern techniques for signal analysis and artificial intelligence. This support tool for decision making accelerates the restoration process, providing greater security, reliability and profitability to utilities. The fault detection procedure is performed using statistical and direct analyses of the current waveforms in the wavelet domain. Current and voltage signal features are extracted using discrete wavelet transform, multi-resolution analysis and energy concept. These behavioural indices correspond to the input vectors of three parallel sets of fuzzy ARTMAP neural networks. The network outcomes are integrated by the Dempster-Shafer theory, giving quantitative information about the diagnosis and its reliability. Tests were carried out using a practical distribution feeder from a Brazilian electric utility, and the results show that the method is efficient with a high level of confidence.