Line loss includes technical and non-technical line losses, which is an important technical indicator for the economic operation of power grid. In response to the problem of limited labeled samples in current line loss anomaly detection, which makes it difficult to determine the location of anomalies, this paper proposes a data hybrid driven line loss anomaly diagnosis method in distribution network, which includes three stages: abnormal feeder detection, abnormal period detection, and abnormal position detection. In the stage of abnormal feeder detection, the first step is to extract abnormal feeder detection features. When there are insufficient labeled samples, clustering algorithms are used for detection. After accumulating sufficient labeled samples, classification algorithms are used for detection to improve accuracy；In the detection stage of abnormal time periods, the X-bar control chart theory is introduced to determine the time periods that exceed the upper and lower limits of control as abnormal time periods；In the abnormal position detection stage, three risk indicators are constructed, and on this basis, a transformer risk level judgment criterion was proposed to locate the abnormal position. Finally, simulation analysis is conducted based on actual operating data to verify the correctness and effectiveness of the proposed method.