The fault diagnosis of gearboxes and bearings in wind turbines is crucial to extend their service life and reduce maintenance costs. This paper proposes a novel fault diagnosis method combining the refined generalized composite multi-scale state joint entropy (RGCMSSJE), robust spectral feature selection (RSFS) unsupervised learning framework, and extreme learning machine (ELM). The method enables feature extraction, dimensionality reduction, and pattern recognition to identify the different condition states of gearboxes. In this method, multi-scale average Euclidean divergence is adopted to assist the RGCMSSJE in parameter selection. Then, the RGCMSSJE is utilized to extract multi-scale features from the gearbox vibration signal and construct a high-dimension feature set. The RSFS method is subsequently used to reduce the dimensionality of the RGCMSSJE feature set. Finally, the reduced low-dimensional features are fed into an ELM classifier for fault pattern recognition. The effectiveness of the proposed fault diagnosis method is verified using two experimental datasets, for which the average accuracy reached 99.9% and 99.3%, respectively. The analysis results show that the method can effectively and accurately identify different fault types in wind turbine gearboxes.