Abstract: At present, the detection efficiency of belt conveyor drum fault detection methods is low, the recognition accuracy is not high, and the feature extraction capability is poor. In order to solve the above problems, a belt conveyor drum fault detection method based on improved YOLOv5s is proposed. A small-sized detection layer has been added to the YOLOv5s network model, making it easier to detect smaller drum faults. The method introduces the convolutional block attention module (CBAM) between the Backbone and Neck to improve the accuracy of target detection. The method introduces efficient channel attention mechanism (ECA) in Neck to enhance feature extraction capabilities for drum faults. The experimental results show the following points. ① On the premise of meeting the real-time detection requirements, the average recognition accuracy of the improved YOLOv5s network model reaches 94.46%, which is 1.65% higher than before the improvement. ② The average accuracy of the improved YOLOv5s network model for detecting drum opening, rubber coating wear, and rubber coating detachment are 95.29%, 96.43%, and 91.65%, respectively, which are 1.56%, 0.89%, and 2.50% higher than before the improvement. A belt conveyor drum fault detection system based on improved YOLOv5s is designed and validated. ① The experimental platform test results show that the average accuracy of the belt conveyor drum fault detection system based on improved YOLOv5s for drum welding, rubber coating wear, and rubber coating detachment detection reach 95.29%, 96.43%, and 91.65%, respectively. The average accuracy of the three types of faults reaches 94.46%, and the detection speed is about 14 frames/s. ② The on-site test results show that the confidence levels for rubber coating wear and rubber coating detachment are 0.92 and 0.97, respectively. The fault type and location of the drum can be accurately identified. This indicates that the improved YOLOv5s-based belt conveyor drum fault detection system is feasible.