Abstract: At present, research on fault diagnosis of mine belt conveyors mainly focuses on three aspects: single-signal detection, traditional algorithm modeling, and multi-feature fusion. Diagnostic methods based on single signals such as vibration and current are prone to problems including bias in feature extraction and insufficient reliability of diagnostic results. Some optimization algorithms suffer from low efficiency in parameter optimization and poor adaptability to multiple fault types. In addition, existing multi-feature fusion studies lack specificity and cannot achieve complementary validation among multidimensional signals. To address these problems, an intelligent fault diagnosis method for mine belt conveyors based on multi-source signal fusion and Bat Algorithm (BA)-optimized Sequential Minimal Optimization (SMO) parameters, namely BA-SMO, was proposed. A vibration–temperature–smoke multi-source signal collaborative acquisition system was constructed. Linear trend removal and an improved Kalman filtering method were used to perform signal denoising preprocessing. An improved Variational Mode Decomposition (VMD) algorithm incorporating an adaptive penalty factor and a redundant component elimination mechanism was proposed and combined with multiscale sample entropy to achieve accurate quantitative extraction of fault features. Based on the extracted multidimensional feature vectors, a BA-SMO model was constructed, in which the global optimization capability of BA was used to optimize the core parameters of SMO, thereby improving the classification accuracy and environmental adaptability of the model. The experimental results showed that: ① the signal-to-noise ratio of the improved VMD algorithm reached 27 dB, and the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) remained below 0.08. The algorithm showed significant advantages in signal decomposition accuracy, efficiency, and matching accuracy of fault feature frequency, and accurately separated the characteristic frequencies of multiple types of faults in mine belt conveyors. ② The BA-SMO model achieved high recognition accuracy for various faults. The recognition accuracy for bearing inner race faults was close to 100%, and the recognition accuracy for idler slip faults was above 90%. ③ Under low, medium, and high interference conditions, the average recognition accuracies of BA-SMO were 99.2%, 97.6%, and 95.3%, respectively. The missed detection rate was below 5%, and the average recognition time was only 32.6 ms. Field application results showed that during a three-month field application, the proposed method successfully identified faults including bearing inner race pitting, idler slip, and rolling element wear. The diagnostic accuracy reached 97.8%, which improved the diagnostic accuracy by 25.3% compared with the traditional manual inspection method and effectively reduced the missed detection and misdiagnosis rates.