The longitudinal tear of conveyor belts is the most common accident occurring at the workplace. Given the limitations on accuracy and stability of current single-modal approaches to detecting the longitudinal tear of conveyor belts, a solution is proposed in this paper through Audio-Visual Fusion. According to this method, a linear CCD camera is used to capture the images of the conveyor belt and a microphone array for the acquisition of sound signals from the operating belt conveyor. Then, the visual data is inputted into an improved Shufflenet_V2 network for classification, while the preprocessed sound signals are subjected to feature extraction and classification using a CNN-LSTM network. Finally, decision fusion is performed in line with Dempster-Shafer theory for image and sound classification. Experimental results show that the method proposed in this paper achieves an accuracy of 97% in tear detection, which is 1.2% and 2.8% higher compared to using images or sound alone, respectively. Apparently, the method proposed in this paper is effective in enhancing the performance of the existing detection methods.