Abstract: To address the high parameter sensitivity of traditional variational mode decomposition (VMD), its tendency toward over-decomposition under strong noise conditions, and the loss of weak fault features caused by hard-threshold screening, an improved adaptive filtering variational mode decomposition (IFVMD) method was proposed in this paper for bearing fault diagnosis. First, a robust power spectrum peak detection strategy combined with a perturbation consistency evaluation strategy was adopted to adaptively determine the optimal number of modes and the initial center frequency of each mode, and differentiated filtering bandwidths were configured for different modes according to spectral flatness. Second, multiple indicators, including modal impulsiveness, periodicity, energy contribution, and spectral structure, were integrated, and a Sigmoid continuous soft-weight function was introduced to complete signal reconstruction, replacing the traditional hard-threshold screening mechanism, thereby achieving adaptive extraction of effective fault modes and suppression of noise interference. Finally, a residual network (ResNet) was used to mine deep discriminative features from vibration signals, enabling intelligent bearing fault identification. Validation results based on the CWRU bearing dataset show that under two load conditions, the average recognition accuracy and recall of the proposed method both reach 99.00%, and the average F1 reaches 99.36%. Under −6 dB strong noise conditions, compared with the traditional VMD-ResNet model, the proposed method improves the recognition accuracy by 5.58% under 0 kW and by 5.15% under 0.746 kW. The proposed method can effectively preserve weak fault features, improve fault identification stability under strong noise environments, and reduce the probability of missed identification of weak faults, thereby providing theoretical support and technical reference for high-precision intelligent fault diagnosis of belt conveyor bearings under complex working conditions.