Condition Monitoring (CM) is considered an effective method to improve the reliability of wind turbines and implement cost-effective maintenance. This paper presents a single hidden-layer feed forward neural network (SLFN), trained using an extreme learning machine (ELM) algorithm, for condition monitoring of wind turbines. Gradient-based algorithms are commonly used to train SLFNs; however, these algorithms are slow and may become trapped in local optima. The use of an ELM algorithm can dramatically reduce learning time and overcome issues associated with local optima. In this paper, the ELM model is optimized using a genetic algorithm. The residual signal obtained by comparing the model and actual output is analyzed using the Mahalanobis distance measure due to its ability to capture correlations among multiple variables. An accumulated Mahalanobis distance value, obtained from a range of components, is used to evaluate the heath of a gearbox, one of the critical subsystems of a wind turbine. Models have been identified from supervisory control and data acquisition (SCADA) data obtained from a working wind farm. The results show that the proposed training method is considerably faster than traditional techniques, and the proposed method can efficiently identify faults and the health condition of the gearbox in wind turbines.