The invention provides a belt conveyor fault diagnosis method based on sound signals, which can reduce the labor intensity of inspection personnel and has the characteristics of high detection speed, high real-time performance, high safety and the like. The diagnosis method comprises the following steps: S1, collecting a sound signal of the belt conveyor; S2, carrying out improved wavelet threshold de-noising processing on the collected sound signals; s3, performing MFCC and deep learning feature extraction on the noise-reduced sound signal of the belt conveyor; s4, establishing a support vector machine classification model, and forming a trained SVM model; and S5, putting the extracted feature information data into the trained SVM model to obtain a posterior probability, then carrying out decision-level fusion by utilizing a D-S evidence theory, and finally, matching a fusion output result with the running state of the belt conveyor in the known running state of the SVM, the running state with the highest matching degree with the fusion output result corresponding to the current running state of the belt conveyor, so that the fault diagnosis of the belt conveyor is completed.