The continuous growth of wind energy generating sources, especially in harsh environments such as off-shore, has led to an increasing demand on more careful planning and control of operation and maintenance costs. This has made condition monitoring and fault diagnosis of wind turbines an even higher priority [1, 2, 3, 4, 5]. In this work failure of a turbine or it’s components is defined as being the state of non-operation of aforesaid turbine or components. A fault on the other hand is associated with a defect, e.g. a crack in the bearings. As wind turbines 10 are composed of different systems, it is intuitive that there are sub-systems which are more vulnerable to failure than others: the rotor system including the hub, has a more dominant failure rate than bearing and generator systems. However, the generator, hydraulic and gearboxes anecdotally are considered the bête noires of wind turbines, as, although not that frequent, the downtime caused by failure in these systems is substantial [6, 3, 7]. There are a wide variety of monitoring approaches available, ranging from acoustic analysis to visual inspection. These have shown potential in early fault detection, with prediction horizons ranging from seconds to months before a failure [8, 9, 10, 11, 12, 13, 14, 15, 3, 16]. Amongst these are approaches that specific target bearing monitoring approaches [17, 18]. Common causes for bearing failure are excessive load, fatigue, contamination, misalignments, overheating etc., latter will be addressed in the course of this paper. Additional common prediction, operation, and condition monitoring approaches are summarized in Kusiak et al. [19] and Márquez et al. [20]. As the proposed method in this work is a fault estimation (including prediction) approach, the comparison to other approaches in the field is essentially the comparison between the structure of set approaches. In general, fault estimation can be categorized into two groups: model based, and data-driven. In case of the first group, a physical model (such as Vidal et al. [21]) or at least an approximate state space model of the system (such as Gao et al. [22] and Liu et al. [23]) is necessary. Given initial information of the system and the consistency between the real and estimated variables, these methods have shown to be successful in providing robust fault estimation. The second group, on the other hand, is solely based on the recorded data, particularly suited when no system informations are available. Although some of the data-driven methods employ a system model to generate residuals (as it is the case of this work), the employed models are unsupervised, such that no prior information of the system is used. The proposed approach in this study is data-driven and aims to predict baring fault based on the statistical features of residuals.