Research On Methods Of Intelligent Fault Diagnosis For Wind Turbine Drive Train Based On Unsupervised Learning

Effective condition monitoring and fault diagnosis and a more accurate view of running condition of wind turbine make sense for the protection of safe operation, loss of investment, optimization of maintenance schedule and the reduction of cost for maintenance. Chinese wind power industry has developed rapidly. With the rapid growth of installed wind energy capacity, grave challenges for maintenance and fault diagnosis are posed. To meet the needs of energy internet and big data technology, fault diagnosis technology of wind turbine needs to develop in intelligent direction. Although a mass of research about fault diagnosis based on pattern classification has been proposed, little is applied for the wind turbine fault diagnosis. The main obstacle is the lack of complete training samples. In actual fault diagnosis, it is difficult to obtain complete training samples with a good generality. The intelligent pattern classification methods based on supervised learning with training samples become increasingly difficult to apply. In this background, this paper study intelligent fault diagnosis methods of wind turbine based on the pattern classification of unsupervised learning for improving the efficiency and accuracy. The main research and conclusion are as follows:(1) As the operating gearbox is influenced by its working environment, it is lack of complete training samples. An adaptive resonance theory (ART) neural network, ART-2 neural network based on unsupervised learning is proposed in order to recognize the changing trend of gear operating condition without known training samples. Thus early fault diagnosis is achieved. Wavelet packet transform is studied and the relative wavelet packet energy is extracted as the features for signal samples. Meanwhile, a method for selecting the threshold value automatically is proposed for improving the adaptiveness of ART-2 neural network. The gearbox vibration data from testing platform used for simulating the fatigue failure process of a gearbox is analyzed by the proposed method. The Results show that the propose method can recognize the changing trend from the normal state to a crack fault before the occurrence of a broken tooth fault. And this method is used for analyzing the bearing operating condition on the generator drive end in a wind turbine. The effectiveness of the proposed method is verified.(2) The winning neural of ART-2 neural network is selected by "hard competition". Only the neural with the biggest output is activated and tested by the similarity threshold value. However, the neutrals with second biggest or the third biggest output may accord with the similarity threshold value testing, which may result in wrong classification for the patterns with similar features. To overcome this issue, a method with ART-2 neural network and C-mean clustering algorithm was proposed. On the basis of ART-2 neural network, C-means clustering algorithm was introduced to modify the classification of samples and the problem of low classification accuracy caused by the hard competition of ART-2 was solved. The Simulation data and actual gearbox fault data were analyzed and the results show that the proposed method can classify and recognize the data with different types. A fault diagnosis scheme of wind turbine group with this proposed method is carried out. The monitoring date from the undiagnosed wind turbines is combined with that of normal wind turbines. The mixed data is classified by the proposed method. The condition of undiagnosed wind turbines is judged through the classification results. Result show that the proposed method can recognize the faulted wind turbine.(3) In order to diagnose both the known faults and unknown faults of wind turbine gearbox, a method based on kernel fuzzy c-means clustering and gravitational search was proposed. Firstly, the clustering model was built based on wrong classification rate of training samples. The training samples were classified by kernel fuzzy c-means clustering. Then the gravitational search method was introduced for solving the clustering model. The class centers of optimal clustering result were acquired. Finally, the similarity parameters in kernel space between new data samples and the class centers were calculated for diagnosing whether the new data sample belongs to knows faults.The actual fault data from wind turbine gearboxes is analyzed by the proposed and the analysis result is compared with that of BP neural network and KFCM algorithm. The results show that the proposed method can diagnose both the known faults and unknown faults effectively. Two optimization algorithms are analyzed and discussed for the fault diagnosis scheme, i.e. gravitational search and particle swarm optimization. The result shows that accurate diagnosis results can be obtained by using both of the two optimization algorithms. However, the particle swarm optimization algorithm takes less computing time, which indicates that it may improve the running speed and diagnosis efficiency of fault diagnosis system.