In order to prevent causing network incorrectly incline with one of fault type after training, the number of training samples for every fault should be allocated averagely.