Aiming at a variety of network attacks on smart meters (SM), which are widely used in smart grid, may be subjected to in the process of measuring and monitoring power consumption, a new abnormal pattern detection framework is proposed to prevent energy fraud in smart meters. The proposed method first sends the user’s power characteristic data to the aggregator based on smart meter. And the distributed data model is used to aggregate the data to better solve the problem of user privacy protection. Then deep belief network (DBN) is used to compare the obtained data with the expected data to better obtain the data features and optimize the top-down features of the training results. Finally, the aggregator marks the SMs in clusters from 1 to N, and transmits the execution data to the meter data management system (MDMS), check through the deep belief network extraction features and replaces the faulty or damaged SM, for more accurate non-technical losses detection analysis. The experimental results show that the proposed method has higher detection rate and applicability than the traditional smart meter data anomaly detection.