Abstract: To address the issues of low accuracy and poor real-time performance in distribution network fault diagnosis, this paper proposes a fault diagnosis method based on the lightweight dual branch feature fusion-AlexNet(DBff-AlexNet). firstly,the collected three-phase voltage and current data from each node of the distribution network are converted to the two-dimensional time-frequency images based on the continuous wavelet transform. These images are then input into DBff-AlexNet for fault diagnosis. Secondly,in the proposed DBff module, the fire skip module adopts squeeze and expand layers with small convolutional kernels to reduce the number of parameters while extracting local features. To preserve information lost in squeeze layers, skip connections are employed for concatenating input features with features from both squeeze and expand layers. Moreover, depthwiseseparable dilated convolutions with varying dilation rates areproposedto expand the receptive field without increasing the parameter count, thereby enhancing global feature extraction capability. finally, the Grad-CAM heatmap method is used to visually illustrate the model's attention regions in time-frequency diagrams, improving the model's interpretability. The simulation results show that this method has higher fault diagnosis accuracy and effectiveness, ensuring the accuracy of fault diagnosis in distribution networks while reducing model complexity.

Key words: distribution network, fault diagnosis, lightweight, wavelet transform, DBff-AlexNet