Non-intrusive load monitoring is one of the important technologies of the ubiquitous power IoT on the customer side, which not only helps the power company to strengthen load management, but also can guide users to rationally arrange the use of the load. In order to achieve the demand side with household power users as the main body, it provides important technical support for responding to and satisfying the demand of residents for precise and lean electricity service. In this paper, for the problem of low resolution of low-frequency sampling signals in non-intrusive load identification and easy overlap of load characteristics, two confluent transient current waveforms and time-domain characteristics are proposed for convolutional neural networks that cannot effectively identify loads with similar waveform characteristics. One of the improved methods is to integrate the root mean square(RMS) of the transient current value into the current waveform image, and the other is to superimpose the threshold judgment on the basis of the identification result of the convolutional neural network to improve the recognition accuracy of the similar waveform feature load. Through the measured data and reference energy disaggregation data set (REDD) test, the feasibility and effectiveness of the proposed method are verified.