With the application of deep learning models in the field of non-intrusive load monitoring, the ability of load identification and decomposition has been significantly improved. However, most methods still have low training efficiency, insufficient decomposition accuracy and the models are difficult to be generalized. Aiming at the above problems, the non-intrusive load monitoring framework is studied with the convolutional neural network based on the time series classification, and the corresponding load identification and decomposition approach is proposed. Through comparative experiments, it is proved that for the dishwasher in the UK-DALE, the convolutional neural network can improve the recognition accuracy by 4.3% and the precision by 19%, while the mean square error of load decomposition is reduced by 21.3%. For the refrigerator in the REDD, the recognition accuracy, precision and F1 score are all improved. Especially, the recall value has increased by 24.3%. In terms of decomposition performance, the mean square error is reduced by 15.8%. Therefore, compared with other neural network models, the convolutional neural networks based on time series classification has more stable recognition and decomposition performance.