Abstract: With the increasing demand for intelligent fault diagnosis in radar systems and the rapid development of machine learning technology, machine learning methods have been widely applied in the field of rapid fault localization for internal components of radar. Addressing the current issues in radar systems where line-type faults cannot be directly fed back through the system, requiring manual troubleshooting which leads to inefficiency, as well as the difficulties in detecting small-sized and complexly wired network cable defects in diverse environments, a radar ethernet cable defect detection model—YOLOv8s-EWD—is proposed. First, the HA_C2f module in the proposed model enhances the ability to express local features. Then, during the downsampling process of the head network, a depthwise convolution module combined with the C2f module is used to reduce the number of parameters in the head network, effectively reducing the model′s parameter count while ensuring detection accuracy. Finally, a new P2 detection layer is added to strengthen the ability to capture fine-grained features. Experimental results show that the YOLOv8s-EWD model achieves significant improvements in fine-grained feature recognition, local feature description, model light-weighting, and detection accuracy.