Abstract: Current research on conveyor belt deviation detection mainly focuses on extracting the straight-line features of belt edges. The method requires setting specific thresholds and is easily affected by environmental factors, resulting in slow detection speed and low accuracy. To address the issue, an underground conveyor belt deviation fault detection algorithm based on an improved real-time detection transformer (RT-DETR) was proposed. The improved RT-DETR was used to directly detect a set of idlers and identify deviation based on the exposure degree of the left and right idlers. Three improvements were made to the RT-DETR backbone network: ① To reduce the number of parameters and floating-point operations (FLOPs), FasterNet Block was used to replace the BasicBlock in ResNet34. ② To enhance model accuracy and efficiency, the concept of structural reparameterization was introduced into the FasterNet Block structure. ③ To improve the feature extraction capability of FasterNet Block, an efficient multi-scale attention (EMA) Module was incorporated to capture both global and local feature maps more effectively. To expand the receptive field and capture more effective and comprehensive contextual information for richer feature representation, an improved high-level screening feature fusion pyramid network (HS-FPN) was adopted to optimize multi-scale feature fusion. Experimental results showed that compared to the baseline model, the improved RT-DETR reduced parameters and FLOPs by 8.4×10⁶ and 17.8 G, respectively. The mAP@0.5 reached 94.5%, with a severe deviation detection accuracy of 99.2% and a detection speed of 41.0 frame per second, outperforming TOOD and ATSS object detection models, meeting the real-time and accuracy requirements of coal mine production.