Smart meters are the basic unit of the energy metering system, which are widely deployed on the user side. Aiming at the problem of the difficulty of maintenance due to their large amount, a smart meter reliability estimation model based on multi-source big data fusion is proposed in this paper. In order to fully explore the useful information in the design, maintenance and operation data of the smart meter, the multi-source big data is merged and collated to obtain the covariate data and the smart meter survival label that affect the life span of smart meters. Based on the survival analysis theory, a cox proportional-hazards（CoxPH） model for the life cycle of a smart meter is established, and a deep neural network is used to characterize strongly non-linear correlation parameters to form a reliability estimation model for the smart meter. Based on the actual operation and maintenance data of smart meters of some city, the effectiveness of the proposed model is verified. Test results show that the proposed model, which provides subsidiarity for smart meter maintenance, and can estimate reliability of the smart meter successfully based on its real-time running status.