The invention discloses a wind turbine generator main bearing fault diagnosis method containing unknown faults. The method comprises the following steps: carrying out efficient time-frequency decomposition on a vibration signal of a main bearing of a wind turbine generator by adopting K-S conversion; extracting features from a complex number time-frequency matrix obtained through K-S decomposition, extracting ten features including a peak value, a mean value, a standard deviation, a variance, skewness, kurtosis, a root mean square value, a peak-to-peak value, Shannon entropy and Renyi entropy from a high frequency domain, a middle frequency domain and a low frequency domain respectively, and constructing a 30-dimensional original feature set; performing descending sorting on the 30-dimensional features according to the feature Gini importance, and selecting the first 15-dimensional features having the maximum influence on classification to construct an optimal feature subset; and finally, identifying the mechanical state of the main bearing of the wind turbine generator containing the unknown fault by adopting an OCSVM and RF combined hierarchical hybrid classifier. According to the invention, new faults of the main bearing of the wind turbine generator can be well identified, potential safety hazards of the main bearing of the wind turbine generator can be found as early as possible, and the operation reliability of equipment is improved.