The traditional methods of detecting arc faults of power lines have the disadvantages of low accuracy and short monitoring distance. To deal with this issue, we in this paper propose a LG-XGBoost collaborative optimization based arc fault detection approach for power lines in home building. First, 21 dimensional characteristics are screened out as the feature vector for identifying arc fault. Then, logistic regression (LG) is used to dynamically adjust the weights of these 21 features. After that, we use the Extreme Gradient Boosting (XGBoost) algorithm as the nonlinear classifier for arc detection by deeply mining the features of arc fault. Finally, dynamic restraint mechanism with 2% error tolerance threshold is introduced to further lower false alarm rate. The experiments show that the proposed detection algorithm has an F1-score of 96.03% under multiple load conditions, and the detection accuracy is 80% even for 500 m distance.