In the actual home environment, it is difficult to collect fault data for household loads, resulting in the scarcity of fault samples and the inability to meet the training requirements of the fault model. In this paper, an IPCNN series fault arc detection method based on transfer learning was proposed. Firstly, an experimental platform for series arc faults of household loads was built to obtain the one-dimensional voltage signals of inductive loads and resistive loads in series faults, and converted them into two-dimensional images by using the Gragram angle field to form a new image dataset and send it to the PCNN model on the source domain for training to obtain the weight parameters of the model. Then, the trained weight parameters on the source domain are migrated to the IPCNN model on the target domain through transfer learning, which accelerates the model training time and saves computing resources. At the same time, GRU and MSA are added to the IPCNN model to improve the computational efficiency and expressive ability of the model, and the classification layer in the PCNN model is discarded, and the L2-SVM is used instead of the Softmax layer to control the complexity of the classification task in the IPCNN model, so as to improve the generalization ability of the model. Finally, in order to solve the problem that the learning rate and the number of neurons in the model are difficult to determine, the improved artificial lemming algorithm is used to optimize the network structure more reasonable. Through comparative experiments, the average recognition accuracy of the model for inductive and resistive loads is 97% and 97.75%, respectively. It is proved that the proposed method overcomes the problem of low model recognition accuracy in the case of data scarcity, and has good results in the identification of series arc faults of household loads.