Abstract: To address the issues where existing abnormal pipeline detection models in coal mine scenarios suffer from low recognition accuracy due to complex target scale variations, large parameter sizes that cannot meet the deployment requirements of inspection robots, and the inability to accurately detect fallen pipelines, an intelligent detection method for abnormal pipelines in coal mine roadways based on an improved YOLOv8n was proposed. Based on the YOLOv8n model, MobileNetV4 was introduced as the backbone network to reduce the number of parameters and improve feature extraction efficiency. An Inverted Residual Efficient Multi-Scale Attention (iEMA) mechanism was introduced to enhance the ability to capture multi-scale target features. In the neck network, Omni-Dimensional Dynamic Convolution (ODConv) was used to replace standard convolution, thereby improving local fine-grained feature fusion and detection accuracy. The Focal-EIoU loss function was introduced to address the imbalance between small-scale leakage targets and slender pipeline samples and to improve the quality of bounding box regression. For fallen pipeline detection, a dual-indicator method based on the overlap between pipeline and ground anchor boxes and the pipeline inclination angle was proposed. When the pipeline anchor box overlapped with the ground anchor box, the pipeline was judged to have fallen. Otherwise, it was judged not to have fallen, and its inclination angle was compared with a preset threshold. If the inclination angle exceeded the threshold, the pipeline was judged to be hanging down. The results showed that, compared with the YOLOv8n model, the improved YOLOv8n model reduced the number of parameters by 19%, while the precision, recall, and mAP@0.5 increased by 3.3%, 4.7%, and 3.9%, respectively. The improved YOLOv8n model achieved a balance between detection accuracy and speed, with no missed detections in small-scale leakage targets and fallen pipeline detection. After deployment on edge devices, the improved YOLOv8n model accurately detected abnormal pipelines, with an average detection time of 9.1 ms and memory usage of only 98 MiB, meeting the requirements of inspection robots for real-time and accurate detection of abnormal pipelines.