Diagnostics of bearings in presence of strong operating conditions non-stationarity - A procedure of load-dependent features processing with application to wind turbine bearings

Condition monitoring of bearings used in Wind Turbines (WT) is an important issue. In general, bearings diagnostics is a well recognized field of research; however, it is not the case for machines operating under non-stationary load. In the case of varying load/speed, vibration signal generated by rolling element bearings is affected by operation factors, and makes the diagnosis relatively difficult. These difficulties come from the variation of vibration-based diagnostic features caused mostly by load/speed variation (operation factors), low energy of sought-after features, and low signal-to-noise levels. Analysis of the signal from the main bearing is even more difficult due to a very low rotational speed of the main shaft. In the paper, a novel diagnostic approach is proposed for bearings used in wind turbines. As an input data we use parameters obtained from commercial diagnostic system (peak-to-peak and root mean square (RMS) of vibration acceleration, and generator power that is related to the operating conditions). The received data cover the period of several months. The method presented in the paper was triggered by two case studies, which will be presented here: first when the bearing has been replaced due to its failure and the new one has been installed, second when bearing in good condition has significantly changed its condition. Due to serious variability of the mentioned data, a decision making process on the condition of bearings is difficult. Application of classical statistical pattern recognition techniques for "bad condition" and "good condition" data is not sufficient because the probability distribution/density functions (pdf) of features overlap each other (for example probability distribution/density function of peak-to-peak feature for bad and good conditions). It was found that these data are strongly dependent on operating condition (generator power) variation, and there is a need to remove such dependency by suitable data presentation. To achieve it, load susceptibility characteristics (LSCh) presenting as feature - operating condition space has been used. Presented 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/rotation speed. Here it has been proven experimentally for the first time that there are two types of susceptibility characteristics related to the type of a fault. The novelty of the paper also comes from an extension of previous study that is statistical processing of data (linear regression analysis) in moving window in the long time of a turbine operation is used for feature extraction. It is proposed here to use novel features for long term monitoring. It will be shown that parameters of regression analysis can be used as unvarying, and fault sensitive features for decision making.