Condition Monitoring of bearings used in Wind Turbines (WT) is an important issue. In general, bearings diagnostics is well-recognized field; however it is not the case for machines working under non-stationary load. An additional difficulty is that the Main Bearing (MB) discussed here, it is used to support low speed shaft, so dynamic response of MB is not clear as for a high-speed shaft. In the case of varying load/speed a vibration signal acquired from bearings is affected by operation and makes the diagnosis difficult. These difficulties come from the variation of diagnostic features caused mostly by load/speed variation, low energy of sought features and high noise levels. In the paper a novel diagnostic approach is proposed for main rotor bearings used in wind turbines. From a commercial diagnostic system two kind of information have been acquired: peak-to-peak vibration acceleration and generator power that is related to the operating conditions. The received data cover the period of several months, when the bearing has been replaced due to its failure and the new one has been installed. Due to serious variability of the mentioned data, a decision-making regarding the condition of bearings is pretty difficult. Application of classical statistical pattern recognition for data from the period A (bad condition) and the period B (after replacement, good condition) is not sufficient because the probability density functions of features overlap each other (pdf of peak-to-peak feature for bad and good conditions). Proposed approach is based on an idea proposed earlier for planetary gearboxes, i.e. to analyse data for bad/good conditions in two-dimensional space, feature - load. It is shown that the final data presentation is a good basis to the very successful classification of data (i.e. recognition of damaged and undamaged bearings).