Research On Angular-domain Vibration Signal Processing And Fault Diagnosis Methods For Wind Turbine Gearbox

Due to the volatility of wind force and the uncertainty of wind direction, vibration signals from wind turbine gearbox under varying rotational speed conditions show typical non-stationarity, whose processing and analysis is significant for wind turbine gearbox condition monitoring and fault diagnosis. Aiming at the problem that there are great difficulties to carry out fault damage test study on actual wind turbine gearbox, a power scaled test bed with similar structure and function to wind turbine gearbox has been constructed under the support of the National Natural Science Foundation of China (Grant No.51375433), so as to explore efficient methods for wind turbine gearbox non-stationary vibration signal processing and fault diagnosis. In the test bed, the one parallel shaft gearbox simulating the high speed stage of wind turbine gearbox was selected and experimented, whose angular domain vibration signal with reference to the output shaft angular displacement was analyzed by order spectrum. Also by using S transform, the approaches of identifying and extracting the periodic impact features from the angular domain vibration signal was presented. Moreover, in later and early life of wind turbine gearbox, two intelligent fault diagnosis methods, namely, radial basis function (RBF) neural network and support vector machine (SVM) were respectively employed and researched, whose focuses were fault feature vectors extracting and diagnosis model parameters optimizing with the use of genetic algorithm.The computed order analysis method was applied to resample the time domain vibration signal with equal angular intervals, which mainly involved resample time points computation and the corresponding amplitudes interpolation calculation. Then by controlling the parallel shaft gearbox rotational speed to vary in a sine curve, the angular domain vibration signals from the gearbox under normal condition and four kinds of fault conditions which are, respectively, small gear with a worn tooth, small gear with a broken tooth, big gear with a worn tooth and big gear with a broken tooth, were analyzed by order spectrum.An intelligent fault diagnosis method using RBF neural network for wind turbine gearbox in later life was proposed. To determine the node centers of the network hidden layer, K-means clustering algorithm was firstly utilized. Then in the experimental research, the fault feature parameters both in angular domain and order domain were extracted from angular domain vibration signals of the gearbox under normal condition and the preceding four kinds of fault conditions. These parameters could be able to adequately reflect the condition information of the parallel shaft gearbox. As train sample and test sample, the fault feature vectors composed of the parameters were input into the network, so as to train the network and test its classification performance. At last, in order to improve the network classification accuracy, genetic algorithm based on real coding was employed to optimize the hidden layer node centers of the network.S transform was introduced to realize spectrum analysis for the angular domain vibration signals from wind turbine gearbox in fault conditions, aiming at detecting the periodic impact features. To restrain noise and highlight the impact features in the angular domain vibration signals, the S transform spectrum was processed by the arithmetic average algorithm and the geometric average algorithm, respectively.For the purpose of extracting periodic impact features from the angular domain vibration signals of damged wind turbine gearbox, two techniques for denoising S transform spectrum were proposed. During singular value decomposition (SVD) denoising process, the target data matrix was composed of S transform spectrum coefficients. The position of the threshold singular value, be less than or equal to which the singular value was set zero, could be determined by the last peak index of the peaks swarm in singular value difference spectrum. Finally, inverse S transform over the denoised data matrix was conducted to reconstruct the angular domain periodic impact features. In the process of denoising S transfrom spectrum by coefficients shrinkage, the optimal threshold was estimated by step iterative algorithm. Then the coefficients were shrunk by hard or soft threshold function according to their modulus magnitude. Similarly, with inverse S trasnform the denoised spectrum could be transformed into angular domain to reconstruct the periodic impact features.An intelligent fault diagnosis method for wind turbine gearbox in early life by utilizing SVM was proposed. In the experimental research, with the use of S transform spectrum arithmetic average algorithm, five standard S transform spectra for angular domain vibration signals of the parallel shaft gearbox under normal condition and the aforementioned four kinds of fault conditons were firstly achieved. Then the fault feature vectors could consist of the cosine similarities and the correlation coefficients between the S transform spectrum of angular domain vibration signal under every condition and these five standard spectra. Meanwhile, five two-class SVMs were constructed with the use of one-versus-all approach, so as to form a five-class SVM. The train sample and the test sample which were made up of the fault feature vectors were input into the five-class SVM to train the SVM and examine its classification performance, respectively. Finally, to improve classification accuracy of the five-class SVM and make its generalization ability better, both the penalty factor and the width of RBF kernel for each two-class SVM were optimized by using genetic algorithm.