Despite their prior knowledge， a large amount of unsupervised industrial data cannot be effectively exploited in real-world applications regions. In this paper， we propose a semi-supervised learning algorithm based on similarity measurement. This method combines specific prior knowledge and unsupervised learning to calculate the similarity between unlabeled data and a small amount of labeled data to augment the minority samples. By improving the classification effect of the model on the minority class， we can mitigate sample imbalance in training phrases and marking difficulty. Empirically， a good minority recognition effect has been achieved in a series of power-sensor detection classification tasks. Compared with state-of-the-art methods， including the traditional and semi-supervised methods， the recall rate and the F₁ value comprehensively exceed the traditional ones.