A gearbox is an important but fault-prone part of wind turbine. Early faults threaten the operational safety of the system. Efficient and accurate gearbox fault diagnosis in extreme conditions is very important for safe and stable operation, thus, a planetary gearbox fault diagnosis algorithm based on an improved deep forest is proposed. To realize the efficient diagnosis of planetary gear box faults in extreme conditions of unbalanced small samples and strong noise, this paper first introduces gradient punishment in a Wasserstein generation countermeasure network to generate samples to supplement the original data set in view of the small and unbalanced vibration data samples of rotating machinery. Second, multi granularity scanning is used to deal with the relationship between vibration signal data points to enhance the fault characteristics in the data. Finally, a new base learner is introduced into the cascade forest and the quantum particle swarm optimization algorithm is used to optimize the key parameters to obtain a model structure with high diagnostic accuracy for fault classification and to output diagnostic results. Compared with other intelligent diagnosis methods in multiple scenarios, it is proved that the proposed method has good classification effect in extreme conditions, and can effectively improve the accuracy of gearbox fault diagnosis.