<p>Accurate fault detection in power transmission systems is critical to maintaining electrical grid reliability. However, existing object detection models struggle in real-world UAV inspection scenarios due to complex backgrounds, scale variability, and irregular target shapes-particularly affecting the detection of small objects. To address these challenges, an enhanced fault detection method based on YOLOv8 is proposed. Firstly, a deformable attention transformer is integrated into the backbone network to dynamically assign feature weights to suppress background interference; Secondly, an auxiliary detection network is designed and an adaptive kernel convolution is introduced to enhance the feature extraction capability of irregular targets; finally, the effective integration of multi-scale features is achieved by linear fusion layer, which improves the utilisation of spatial information and enhances the detection of small targets. The experimental results show that the model achieves precision, recall, and mAP values of 89.55%, 84.54%, and 88.06%, respectively, on PTL-Al_Furnas dataset. Compared with the current state-of-the-art method, the proposed approach improves precision by 2.25% and recall by 4.36%. It significantly improves the performance of fault recognition in complex scenes. Additionally, the model achieves a validation speed of 3.8 ms, fulfilling the requirements for real-time detection, and exhibits robust generalisation across both the transmission line obstacle dataset and substation defect datasets.</p>