This paper proposes a new approach to fault diagnosis in electrical power systems, which presents an aspect little explored in the literature that is the protective device failure detection together with the fault section estimation, since the majority of the methodologies so far proposed to fault diagnosis are limited to the fault section estimation alone. The proposed methodology makes use of operation states of protective devices as well as information related to the protection philosophy. Initially, these data undergo a preprocessing step to convert the format of 0 and 1 to percentage values. The conversion to percentage values allows the use of artificial neural networks, whose numbers of inputs do not depend on the number of alarms of the protection philosophy, or the type of bus arrangement or the number of circuit breakers. This allows the same set of neural networks to be trained and applied in different power systems with different protection schemes and bus arrangements. The proposed system has five neural networks, each containing few neurons and requiring 30 μs to perform fault diagnosis. The proposed system was trained considering the IEEE 57-bus system, containing different selective protection schemes, and subsequently tested in the IEEE 14-bus, 30-bus, and 118-bus systems, and Eletronorte 230-kV real power system.