Abstract: With the increasing demand for energy management of devices by residential users, the non-intrusive load monitoring（NILM） technology with low hardware cost has great engineering application value. At present, the load monitoring deep network has the problem of outputting false positive power segments, which leads to the misjudgment of on/off operation status of the electric appliance and affects the accuracy of load decomposition. Therefore, an NILM method based on multi-task learning is proposed. First, a load monitoring inference framework based on the multi-branch convolutional network and improved postprocessing is established. Based on the main branch of the load power decomposition network, an on/off status evaluation subbranch network is set up, and the supervision information of on/off sequence of the electric appliance is introduced. The gradient backpropagation process during network training is adjusted by the multi-task aggregation loss function, which reduces the misjudgment rate of the operation status. Then, the weighted mean filtering is used to post-process the on/off status evaluation value and power output value of the network to further reduce the load decomposition error. Finally, the comparative experiments are carried out on the UK-DALE data set. The experimental results show that the proposed method can well separate the electric appliance load containing complex power characteristics, and the effectiveness of the method is verified.