The invention relates to the field of machine learning, and provides an anomaly recognition method and device based on semi-supervised deep learning, and a storage medium, and the method comprises thesteps: S110, obtaining sample data; s120, acquiring positive sample data enhancement, negative sample data enhancement and data noise; s130, forming a corresponding annotation data positive sample, an annotation data negative sample and an annotation data noise sample; s140, forming three corresponding initial prediction models; s150, respectively inputting the unlabeled sample data into the three trained initial prediction models for data prediction; s160, labeling the unlabeled sample data to form new labeled sample data; s170, adding new labeled sample data into the initial labeled sampledata, and circularly executing the steps S120 to S170 to form a final prediction model; and S180, inputting to-be-identified data into the final prediction model to perform anomaly identification. According to the method, the requirement for data is low, a large amount of marking data is not needed, and meanwhile the data exception recognition accuracy can be improved.