Abstract: Axial flow threshing and separating device can be one of the most important components in the grain harvester. However, the performance of the device is often confined to the blocking tendency state. In this study, a test bench for the blockage tendency was developed for the axial flow threshing and separating device. The diagnostic models of the blockage tendency were also established for the early warning of the blockage. A series of experiments were conducted to explore the effect of the material feed rate on the blockage tendency. The signals of the drum speed and the shell vibration were collected from the threshing and separating device. One speed and four acceleration sensors were utilized to determine the variation in the drum speed and shell vibration. The results indicate that there were no-load, normal, slight, and severe blockage tendencies. Specifically, a slight risk of the blockage was observed at the normal feeding rate of 0.5 kg/s, as the feeding rate increased by 20%. Furthermore, the duration of the slight blockage tendency suddenly decreased to less than 1 s when the feeding rate increased by 40%. There was a severe risk of the blockage tendency within 0.5 s. The signal adjustment also caused the drum speed with the lag and low accuracy. By contrast, the vibration signals were rich in sensitive information for the early and accurate warning of the blockage. Therefore, the Slime Mold algorithm was used to optimize the variational mode decomposition. The optimal parameters were obtained, including the mode number (K) and penalty coefficient (a). Then, the vibration signal of the sensor was decomposed into multiple intrinsic mode components (IMCs). The speed and vibration signals were verified to divide the operating intervals for each modal component under different working conditions. Nine features were extracted for each state interval, including the center frequency, Kurtosis, energy entropy, fuzzy entropy, approximate entropy, envelope entropy, envelope spectrum mean, and peak. Features 1–9 of all sensors were fused for the feature matrix (583 rows, 9 columns) and the state label matrix (583 rows, 1 column). A weighted mean of vectors algorithm was improved to reduce and then filter 9 features, in order to improve the diagnostic speed. The minimum weighted mean impacts of 233 870.46 in the features 6, 3, 4, and 7 differed from those of 53 370.36 in the features 1–9, which was the minimum of 70.18%. While the minimum weighted mean impacts of 359 675.71 in features 3, 6, 4, and 7 differed from those of 179 663.72, which was the minimum of 50.05%. The weighted mean impacts were better distinguished from the features with the high impact values. The high-sensitivity feature screening was performed to reduce the feature dimensionality. Finally, a blockage tendency of the diagnosis model was combined with the BP neural network. The feature matrix was selected as the input, while the state label matrix was selected as the output. The goodness-of-fit (R²) of the diagnostic model was greater than 0.9. The output was highly correlated with the true labels, indicating the high model reliability. The diagnostic accuracy was 92.57% under different working conditions, with a prediction speed of about 614 samples per second. The blockage tendency of the diagnosis model shared high accuracy and short time. The early warning of the blockage can be used to prevent the blockages for the high operational efficiency of the machine.