Abstract: Automatic verification system of smart meter may be abnormal in the long-term operation process，

but the regular manual detection method can not timely learn the risk information，shortening the period of manual inspection will reduce the work efficiency of automatic verification. In most cases，the sample data of abnormal epitopes are unmarked，and the unsupervised anomaly detection algorithm is usually used to screen abnormal epitopes. In order to reduce the false positive rate of unsupervised anomaly detection and the cost of manual inspection，a request for manual inspection of unsupervised screening "abnormal epitopes" was proposed，in the process of eliminating epitope faults，a small number of labeled samples were obtained，and semi-supervised transductive support vector machine （TSVM） anomaly detection model was constructed by using labeled and unlabeled samples. In the future automatic verification work process，new labeled samples and unlabeled samples were obtained continuously，the TSVMmodel could be extended and optimized according to semi-supervised training method. Using the proposed method，the automatic verification data of state grid Shanghai electric power company was analyzed，and compared with the manual inspection results，the effectiveness of the method was verified.

      Key words: smart meter; automatic verification; verification data; anomaly detection; semi-supervised; TSVM

mode