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.

• 为了防止人为的造成训练后的网络过多的倾向于某一故障类型，各故障类型的训练样本数量不应相差太多。