Aiming at the problems that the faults in low-voltage centralized meter reading system are complex and the current operation and maintenance is difficult to meet the harsh user demand, we proposed a fault diagnosis method for LV centralized meter reading system based on topology analysis-deep learning. Starting from the two stages, planning and operation, we analyzed the transformer-concentrator association and concentrator-electric energy meter association to diagnose the physical topology of LV centralized meter reading system. Based on the determined physical topology and information flow path, a deep belief network fault diagnosis model is automatically established by offline learning with emerging fault events. Online obtaining the vital systematic operation character, the system fault section feature vector is established and sent to the well-trained fault diagnosis model for final diagnosis result. The result of the case study have showed that the proposed method can effectively and accurately diagnose the fault in LV centralized meter reading system, and it’s effective to deal with the case of the missing information and wrong information.