<p>The advancement of distribution line fault diagnosis technology can significantly enhance fault handling efficiency and accuracy, reducing the impact on the power system and ensuring a reliable power supply. These technological advancements have important theoretical and practical implications for improving the operational excellence of the power system. Recent developments in machine learning have sparked interest in machine learning-based approaches for line fault classification. Although these methods show promise in improving classification accuracy, they often rely on a single fault feature type and require extensive simulated data for training. In practical scenarios, the limited availability of line fault samples and the complexity and noise in real-world data pose challenges to the prediction accuracy of machine learning-based methods. In this paper, we establish a systematic framework for diagnosing distribution line faults, effectively addressing complex fault diagnosis scenarios. We extract multi-dimensional temporal-frequency domain features and propose a Dempster–Shafer evidence theory-based multi-model fusion method to significantly improve the accuracy and efficiency of fault identification. Real-world data collection and experiments demonstrate the performance of our approach, achieving an average accuracy of 74.6% in identifying low-resistance grounding, arcing grounding, high-resistance grounding, and intermittent arcing grounding.</p>