Research On Methods Of Fault Diagnosis For Wind Turbine Transmission System Based On Manifold Learning

As a clean and renewable resource, wind energy has been one of the areas that countries in the world are scrambling to develop. In China, wind power generation equipment has achieved rapid development and has formed a new industry. However, due to poor working condition, unstable wind speed and alternating loads, wind turbine transmission system is most likely to fail or damage. Imbalance, wear, fatigue damage and fracture of transmission system are the main failure modes of wind turbine. In addition, the wind turbines that are installed in remote areas are far away from the ground, this causes the wind turbines maintenance, in this case, intelligent fault diagnosis for wind turbine transmission system seem to be very necessary and urgent. But on the other hand, the wind power industry currently existes technology gaps in intelligent fault diagnosis. For example, the local controller SCADA of large-scale wind turbine only runs data acquisition, thresholds alarm and communications functions, and can not drive for intelligent fault diagnosis of wind turbine transmission system. Similarly, wind turbine condition monitoring system WindCon developed by SKF has the early warning, alarm and simple signal analysis functions,but lacks accurate fault diagnosis methods. Meanwhile, the current prevailing model"feature extraction→pattern recognition"of fault diagnosis for rotating machinery exists the following two serious disadvantages that is incompatible with the technology need of fault diagnosis for wind turbine transmission system:①feature extraction methods require manual intervention so that selection quality of fault feature all depends on experience, this method conflicts with automatic, high precise and rapid fault diagnosis way that is required by working conditions and operation mode of wind turbine.②the single or single domain feature extraction approaches widely adopted at present are very difficult in comprehensively characterizing different fault parts, different fault types and different fault severities of complex wind turbine transmission system under complex conditions, but as is well known, wind turbine transmission system is a rigid-flexible coupling unity, thus strong coupling effect among wind wheel, spindle and gearbox of transmission system easily cause the complexity of failure mode and failure hierarchy. This requires that fault diagnosis method of wind turbine transmission system should be promoted to the systematic one. However, the existing fault feature extraction methods are ver difficult to meet the technical requirement. Therefore, based on manifold learning dimension reduction theory, intelligent fault diagnosis methods characterized by their unification of stability, automation, high-precision, rapidity and generality have been studied in this paper. The main research work and conclusions are as follows:①For the drawback that the adaptive pattern recognition characteristic, namely, the stability of pattern recognition accuracy of the current pattern recognition methods is very poor, based on Least Square Support Vector Machine (LS-SVM), two types of solutions are proposed in this paper. First, on the premise of considering the LS-SVM's training model as the objective function of optimization problem, Genetic Algorithm(GA) is adopted to hierarchically optimize the index parameter, compensation parameter and penalty factor of LS-SVM for minimizing objective function. By this way, the improvement of LS-SVM's adaptability can be achieved. Second, wavelet kernels Littlewood-Paley,Shannon and Morlet which can approximate the training objective function and decision function of LS-SVM, i.e. have better nonlinear mapping ability are constructed to improve the adaptive classification property of LS-SVM. In addition, K-nearest neighbors classifier (KNNC), which directly use the local class label information of training samples for classification of test samples, is studied in this paper. Theoretical analysis, simulation experiment and fault diagnosis example of wind turbine bearings show that as for diagnostic accuracy and stability, Wavelet support vector machines LPWSVM,SWSVM,MWSVM and KNNC are equivalent to LS-SVM optimized hierarchically by Genetic Algorithm, but with regard to computational efficiency, the former is better than the latter. Above research work provides improved, efficient and stable pattern recognition methods for wind turbine fault diagnosis.②Facing on the problem that the current prevailing model"feature extraction→pattern recognition"of fault diagnosis, which is hard to realize the combination of automation and high-precision, cannot adjust to technology requirement of fault diagnosis for wind turbine transmission system, a novel fault diagnosis model for wind turbine is proposed based on dimension reduction with Principal Component Analysis (PCA) and Back-propagation(BP) neural network in this paper. In the diagnosis model, Singular value spectrum of original signals'bispectrum correlative character matrix, which acts as fault feature, is firstly solved, after that, High-dimensional Singular value spectrum is rapidly compressed into low-dimensional datas by PCA, then nonlinear processing of PCA is enhanced by BP neural network, finally, the low resolution which BP neural network cause on account of relapsing into local minimum﹑insufficient training or over training is calibrated by SWSVM. Thus, the higher recognition accuracy and adaptive diagnosis capacity can be obtained. As showed by the results of theoretical analysis, simulation experiment and fault diagnosis example of wind turbine bearings, the fault diagnosis model not only overcomes the difficulty in applying bispectrum to wind turbine fault diagnosis, but also realizes the automation and high-precision diagnosis of wind turbine failures.③On the premise that automatic and accurate fault diagnosis of wind turbine transmission system is ensured, a novel manifold learning-based intelligent fault diagnosis mode " mixed-domain feature fusion→Feature compression with manifold learning→pattern recognition"is proposed in this paper to further realize the rapidity and generality of fault diagnosis. Then based on the proposed fault diagnosis mode, a new intelligent fault diagnosis model"the integration of fault feature in time domain and frequency domain/the fusion of EMD and AR model coefficients—Orthogonal locality preserving projection(OLPP)/Orthogonal Neighborhood Preserving Embedding(ONPE)/Linear local tangent space alignment(LLTSA)—wavelet support vector machine(WSVM)/KNNC"is builded. The theoretical analysis, simulation experiment and fault diagnosis example of wind turbine bearings indicate that the builded fault diagnosis model for wind turbine transmission system combines the strengths of mixed-domain feature fusion in extensive extraction of fault feature, manifold learning in effective compression of fault information and artificial intelligence in pattern recognition, and truly realizes the automation, high-precision, rapidity and generality of wind turbine fault diagnosis method. The research on manifold learning, mixed-domain feature fusion and the novel intelligent fault diagnosis model enriches and develops the theories and technologies of wind turbine fault diagnosis.④With intelligent fault diagnosis theoretical models as the leading role, condition monitoring and fault diagnosis system of wind turbine is developed. The condition monitoring and fault diagnosis system, which includes multi-agent components such as data acquisition, signal transmission, database storage management, remote data access, condition monitoring, intelligent fault diagnosis and man-computer interaction, is able to achieve the system integration of condition monitoring and fault diagnosis technologies of wind turbine.At the end of the thesis, the summarization of the article and expectation of the relative technology development are presented.