Abstract: A fusion diagnosis method based on the starfish optimization algorithm (SFOA) and convolutional neural network (CNN) is proposed to address issues of insufficient feature extraction accuracy and high noise sensitivity in 10 kV distribution network fault diagnosis. Four different fault types were selected as samples: single-phase-to-ground fault, short-circuit fault, line break fault, and lightning strike fault. To optimize fault data according to the voltage characteristics of 10 kV distribution systems, two preprocessing methods were applied: Synchronous extraction transform combined with wavelet packet decomposition (SET-WPD), and discrete fréchet distance combined with fuzzy C-means clustering (DFD-FCM). The SFOA-CNN model was then trained and used for fault recognition. By comparing prediction performance on the test set and comprehensively evaluating key algorithmic metrics, the optimal model for 10 kV distribution network fault diagnosis was determined. Results show that the DFD-FCM-SFOA-CNN model performs best, with a goodness of fit of 0.99932, making it the most suitable model identified in this study. This approach significantly enhances fault diagnosis efficiency and contributes to the secure operation of power systems.