Fault Space Dimensionality Reduction For Wind Power Converter Based On Manifold Learning

In the context of the energy crisis and environmental protection, the development and utilization of new clean and renewable energy increasingly pay close attention in society. Wind energy is currently the fastest growing clean and renewable energy, Huge reserves, in the form of a simple conversion and large-scale mining and pollution-free characteristics of wind power, it has become the fastest growing new energy, the most mature technology generation technology. Wind turbines often work in the field, harsh environmental conditions, converter is a core component of wind power systems, so its stable and reliable operation is critical. When the converter fails, if you can not get timely diagnosis and recovery, the entire wind turbine might suffer serious damage or lead to downtime, which brings huge economic losses. Therefore, the reliability requirements for converter of the wind power systems is very high, converter fault diagnosis is helpful for improving system stability.The existence of a large number of power electronics in converters, and for the nonlinear, it is difficult to establish and improve the analytical model, the converter has a wide range of the failure. So the troubleshooting of a complex system can not be pre-existing diagnostic technology to meet needs. Also, the presence of the feature extraction of the current fault diagnosis mode "feature extraction → pattern recognition" requires manual intervention, When the characteristic dimension is high, feature amount has a drawback of redundancy. And troubleshooting indicators for automation, high precision, fast deviate from the required. Therefore, the shortcomings of non-stationary nonlinear characteristic of wind power converter fault signal and difficulty for pattern recognition for high feature space dimension in this thesis, Combine empirical mode decomposition to deal with the nonlinear signal for being adaptive, with that fractal box dimension can quantitatively describe the nonlinear behavior characteristics. Propose a smart wind power converter fault diagnosis named "feature extraction→dimensionality reduction→pattern recognition" based on empirical mode decomposition and box dimension and manifold learning.Firstly, introduce the structure and working principle of wind power converter, and analysis the open fault of IGBT in the converter. To obtain the waveform for normal operation and fault by building Matlab simulation model of the converter, and then select the three-phase output voltage to extract fault features.Secondly, choose the three-phase output voltage to extract fault feature,the empirical mode decomposition of the three-phase voltage signal to calculate the information entropy and box dimension of each intrinsic mode, information entropy is energy characteristics of the signal, box dimension is demographic characteristics. The combination of both as a signal converter characteristic quantities, and effectively to achieve fault features and then fully accurately describe the fault condition. By the simulation results, this method is validity to reintroduce manifold learning method for converter fault feature dimensionality reduction.Once again, introduce the relevant theory and classical algorithm of manifold learning, in theory, a manifold learning method support the applicability for converter malfunction space dimensionality reduction. Focusing on some of the flaws of manifold learning methods, conduct to improve the research, based on the local linear embedding, propose supervised incremental quadrature discriminant neighborhood preserving embedding manifold learning algorithms.Finally, verify the accuracy and effectiveness of the proposed method for converter fault identification by Matlab simulations.