Research On Early Fault Prediction Of Wind Turbine Gearbox Based On Subspace Method

With the crisis of energy and environment, the fast development of wind powergeneration field is undoubtedly a positive thing. But the problem of wind farms highmaintenance costs has followed. The longest downtime for wind turbines is due tofailures of the gearbox, and the gearbox is one of the most expensive components inthe wind turbine. So early fault prediction of gearbox is meaningful for prolongingthe stable operation time and reducing maintenance costs.Supported by the Natural Science Foundation Fund Project (NO.51277074), thispaper presents two based on subspace fault prediction algorithms for vibration data ofthe gearbox.(1) The first fault prediction algorithm is based on the stochasticsubspace identification method. Firstly, the gearbox’s stochastic linear state spacemodel is built up by using a large amount of vibration data under normal operationalconditions, and we obtain a set of reference eigenvalues which are related to thecharacteristics of gearbox’s normal state. Then, new eigenvalues are calculated by thereal-time vibration data and we compare these new eigenvalues with the referenceeigenvalues. The gearbox is normal if the errors between new eigenvalues and thereference eigenvalues are very small, if not, it is abnormal. Next we define aroot-mean-square error index to easily compare the eigenvalues, and then we canpredict the running state of gearbox with the defined index and a thresholddetermined by SPC principle.(2) The second fault prediction algorithm is based onthe combined deterministic-stochastic subspace identification method. The algorithmis the deepening and expending of the former. It not only considers the influence ofgearbox fault prediction on the vibration data, but also considers the influence on thevariable rotational speed data. Firstly, the vibration data and rotational speed data areanalyzed to obtain a reference domain that relate to normal state of gearbox. Then,the parameter matrices of the state space equation can be estimated by another set ofdata and we calculate the matrix’s eigenvalues. The gearbox is normal if theseeigenvalues are within the scope of the reference domain, if not, it is abnormal. Wealso define a dispersion index to describe the simulation results quantitatively. Whenthe gearbox is normal the dispersion curve is always lower than the threshold value,and conversely the curve will be higher than the threshold value.Under the MATLAB environment this paper completes the implementation ofthe above algorithms. Then the algorithms are simulated by the normal vibration dataand fault data respectively, and acquire good results. The results show that theproposed methods are correct and effective.