To solve the erroneous judgement of arc-fault recognition problem

cost-sensitive learning theory is introduced and different types of misclassification are given unequal cost.The current data of typical load in normal condition and arc-fault condition is collected by experiments. Feature vectors are extracted according to the time domain and frequency domain characteristics of current signals. MetaCost method is applied to train cost-sensitive neural networks which is used as recognition module of arc-fault.The experimental result shows that cost-sensitive neural networks with reasonable cost matrix values can reduce the number of erroneous judgement.