Methodological Research On Wind Turbine Gearbox Fault Diagnosis Based On Manifold Learning

The development of wind energy has infinite potential, and the world has available wind energy resources is equivalent to 5 times the global power generation when it used for power generation. Wind power is flexible and economical, and the operation of wind farm is simple. However, frequent fault of wind turbine has increased the cost of wind power, high maintenance costs and downtime causes huge economic losses. According to statistics, the fault occurrence rate of gear box is higher and machine down time is the longest when the main components of wind turbine generator is in fault. The research on advanced fault diagnosis which can achieve the real-time and accurate diagnosis has important significance.This dissertation takes a gear box as research object, which structure is composed of two-stage planetary gear and a parallel shaft. The vibration signals are acquired and analyzed by constructing the simulation model of the gear box. On this basis, a fault diagnosis method based on manifold learning theory is researched by the integrated use of machine learning, vibration analysis and time-frequency analysis technique.Firstly, the thesis elaborate the basic theory and several classical algorithms of manifold learning, and discuss some problems in manifold learning, such as the nearest neighbor parameter selection, the intrinsic dimension estimation, supervised learning, the influence of noise and generalization. Then, the dissertation does research on neighbor parameter selection and supervised learning, and an improved algorithm which is manifold learning algorithm of orthogonal discriminant based on random projection is proposed. This algorithm uses manifold distance to choose neighbor points, rather than the Euclidean distance, so effectively avoid the sensitivity of neighbor parameter, the supervised discrimination ability of algorithm is increased by introducing the maximum margin criterion, the mapping from high dimensional space to a low dimensional space is achieved by random projection, and it can be ensure the distance changes are minimal with high probability between projection data. Simulation results show that the improved algorithm is effectiveness.Next, a three dimensional solid model of 1.5MW wind turbine gear box is built in UG software, then, the model is imported into Automatic dynamic analysis of mechanical systems(Adams) and corresponding constraints and attribute information are added. So the virtual prototype model of gear box is acquired, and the accuracy of model is verified through acquired response curves. To obtain fault signals, artificial faults are added to gear box and fault simulation models are established, then the vibration signals of gear box under normal, broken teeth and shaft misalignment are obtained. Finally, the fault diagnosis of gear box is achieved by using the fault diagnosis model based on manifold learning, “vibration signals?manifold learning algorithm of orthogonal discriminant based on random projection?empirical mode decomposition?energy entropy of intrinsic mode function? support vector machine classification”.According to the problems of neighborhood selection and supervised learning, local linear embedding algorithm is improved. On the basis of manifold learning, dimension and noises of vibration signals of gear box are reduced and eliminated, so the first feature extraction is achieved. The second feature extraction is realized by combining empirical mode decomposition and energy entropy, and finally manifold learning fault diagnosis model is applied to the fault diagnosis of wind turbine gearbox.