Abstract: To address the problem that weak bearing fault features in wind turbine drivetrains are easily masked by noise under complex nonlinear disturbances, and that traditional signal processing methods struggle to extract fault information effectively, a Correlation Frequency Graph Spectrum (CFGS) model was proposed for wind turbine bearing fault diagnosis. First, a Laplacian matrix was constructed from frequency-domain amplitude information to map one-dimensional vibration signals into the graph spectrum domain. Second, a migrated frequency-domain feature analysis method and a linear graph-spectrum order separation strategy were developed to bridge the time domain and the graph spectrum domain. Third, a hybrid correntropy index was constructed to quantitatively evaluate the fault features embedded in different graph-spectrum orders. This enabled unsupervised adaptive selection of the optimal reconstruction order and accurate diagnosis of wind turbine bearing faults. Validation was performed using simulated signals and measured inner- and outer-race fault data of wind turbine bearings. The results show that the model effectively suppresses background noise. It highlights periodic fault impulses buried in noise and fully recovers fault characteristic frequencies, harmonics, and modulation sidebands. At a signal-to-noise ratio of −10 dB, the model achieves more reliable fault feature decoupling and more robust diagnostic performance than Fast Kurtogram and GSS-EWT. The proposed method provides a feasible technical route for weak fault diagnosis of wind turbine bearings under complex operating conditions.