Non-Intrusive Load Monitoring (NILM) systems are widely used for energy management, but their accuracy can be limited by the removal of monitoring equipment after the initial training phase. This removal results in a lack of ongoing training data and impairs the ability to improve recognition accuracy. To address this challenge, this research proposes a novel semi-supervised learning approach for NILM systems. The approach is based on a previously proposed temporal convolutional network-conditional random field (TCN-CRF) model. It uses optimal interval filtering and unanimous voting to control the quality of pseudo-labels. The approach identifies high-precision segments in the sequence-to-sequence algorithm as semi-supervised learning targets and generates pseudo-labels, allowing for continuous self-supervised learning and sustained improvement in recognition accuracy. The results show that the proposed method can effectively combine unlabeled data to improve model accuracy, and achieve up to 14% highest improvement in algorithm accuracy over fully self-supervised methods in small sample training.