A fast and effective fault diagnosis system is crucial for ensuring complex mechanical equipment’s safe and reliable operation. Deep learning has shown promising prospects in fault diagnosis applications, but existing algorithms have limitations in frequency analysis and long-time sequence feature encoding, which greatly restricts the practical application of deep models in the diagnosis field. This article proposes a transformer framework based on fast Fourier transform (FFT), called FFT-Trans, for mechanical fault diagnosis to overcome these limitations. FFT-Trans creatively extends the global information interaction mechanism of the transformer from the time domain to the frequency domain, thereby realizing global correlation encoding in the frequency domain and mining hidden fault features. Specifically, we replace the self-attention layer in the transformer with the global frequency encoding layer (GFE-Layer) and use learnable filters for global information exchange and better multiscale fusion. This approach can transform different types of signals into frequency components for analysis. By analyzing different frequency components in the frequency domain, the fault type and location appearing in the signal can be more accurately determined. In addition, it can fully extract the inherent connection between the vibration signal and the fault, achieving more comprehensive fault detection. We conducted experiments on the high-speed aviation bearings dataset and motor bearing dataset to validate the proposed method. The experimental results show that FFT-Trans has a better performance compared to existing deep diagnostic models, and still has considerable fault diagnosis performance in noisy environments.