Abstract: The traditional operation and maintenance methods for distribution networks encounter difficulties in integrating multi-source heterogeneous data when dealing with large-scale multi-source data. These problems include the difficulty in multi-source heterogeneous data fusion caused by the dimensionality disaster, insufficient accuracy of fault prediction, and unreasonable allocation of maintenance resources. To address these issues, this article proposes a distribution network operation and maintenance method based on big data technology. By utilizing multi-source data such as IoT sensors, production management systems, and meteorological environment, the method achieves the fusion and feature extraction of heterogeneous data through edge computing gateways and D-S evidence theory, solving the problems of data heterogeneity and conflicts. A fault prediction model is constructed by combining one-dimensional convolutional neural networks with the attention mechanism to accurately identify the health status of equipment and reduce the risk of false alarms. An integer programming model is used to optimize differentiated maintenance strategies, achieving hierarchical management of maintenance tasks under resource constraints. Experiments show that the false detection rate and missed detection rate of this method are lower than those of existing methods, indicating that this method significantly improves the accuracy and resource utilization efficiency of distribution network maintenance.