Topology recognition of power distribution networks is the foundation for the planning, operation, and scheduling of power grids. The introduction of neural network can improve the accuracy of topology recognition in power distribution networks, but how to effectively design the neural network model that can balance robustness and convergence is still a difficult problem that needs to be solved urgently. In this article a method for topology recognition of power distribution network based on data fusion with multi-channel and multi-scale convolutional neural networks is proposed, where data such as voltage amplitude, voltage phase angle, active power and reactive power of each node in a power distribution network are preprocessed and input into independent channels; in each channel the input data is subjected to convolution and normalization calculation at multiple scales to extract multi-scale convolution features, these normalized features are fused through a 1 × 1 convolution kernel and then mixed with the original input data of the channel, the mixed data are then processed through nonlinear transformation and output; the above data processing processes of each channel are carried out in parallel, the data output from all channels are flattened and input into a topology recognizer which is composed of fully connected layer network and a Sigmoid layer to complete the task of topology recognition of power distribution network. This proposed method is characterized by its scalable model and less affected by the quality of the original power data. Through case study on the power distribution networks of IEEE 33, it has been verified that compared to the 1D-CNN, Resnet18 and DNN methods, the proposed method has the advantages of fast convergence speed, strong anti-interference performance, high robustness. The study demonstrates that the proposed method, based on reasonable feature extraction and fusion utilization of various power data, effectively addresses the challenge of balancing robustness and convergence in topology identification models. It reduces the model training''s reliance on large-scale datasets and is applicable to distribution network scenarios with distributed power sources, providing a critical foundation for distribution network system analysis and scheduling management.