Aiming at the polishing machines for grain processing in the actual operating environment of the signal complexity and non-stationarity and other issues, this research proposes a fault diagnosis method based on Deep Temporal Feature Fusion Network (DSTFN-VMDNR). The process begins with Variational Mode Decomposition (VMD), which denoises the original signal and extracts multiple stationary modal components. Deep temporal features and multi-scale features are then extracted using the Time-Series Network (TimesNet) combined with a Bi-directional Long Short-Term Memory (BiLSTM) network. A convolutional residual block is incorporated to further enhance feature extraction capability. Finally, the extracted features are aggregated using a Global Average Pooling (GAP) layer. The model achieves accuracies of 99.75% on the CWRU dataset and 96.87% on the Jiangnan University dataset, thereby demonstrating the effectiveness of the method for bearing fault diagnosis.