In view of the problem that the early fault signals of bearings in industrial equipment are weak, easily disturbed by environmental noise, and the traditional fault diagnosis methods is difficult to capture the fault characteristics, this paper proposes an early warning method that integrates frequency domain vibration analysis and improved Transformer. First, a multi-scale frequency band decomposition network is constructed to extract sensitive frequency band features in vibration signals through adaptive wavelet packet transform. Then, a dual path feature fusion architecture was designed. Finally, a dynamic weight allocation strategy was proposed. Experiments show that the improved Transformer can achieve an accuracy of 98.7% in 5-h early warning; under different signal-to-noise ratio environments, especially under noise conditions below − 5 dB, the model can still maintain a high detection accuracy. Dynamic weight allocation strategy significantly improves overall performance, especially under − 10 dB noise conditions. The experimental results demonstrate that the synergistic effect of integrating frequency domain features and improved timing modeling can detect faults in early weak fault detection.