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

Wind turbines operate under hostile conditions such as variable operating load and rotating speed, high temperature difference, and so on. As a result, components and parts of wind turbines are prone to failure. Especially the transmission system always suffers from failure, which leads to long time shutdown of wind turbine. The structure of transmission system is complex with large size, and as a result installation and maintenance of wind turbine transmission system is very difficult and repair cost is high. What is worse is that the long time shutdown caused by transmission system faults seriously affects the power generation and result in economic loss. Incipient fault diagnosis is the key to prevent the huge economic losses caused by wind turbine transmission system faults, and that is to say the fault should be accurately diagnosed when the fault is in the embryonic stage or the fault is slight. The variable operating conditions, hostile working environment and complex structure of wind turbine transmission system lead to that the fault features of incipient transmission system fault are weak and submerged in strong environmental noise, so fault features is hard to extract; Meanwhile, vibration monitoring of wind turbine transmission system has just begun, and there lacks systematic, long-term collection of fault samples of incipient transmission system fault, which result in that the fault sample is scarce. The above characteristics of the wind turbine transmission systems lead to that accurate diagnosis of incipient transmission system faults beyond the capacity of the traditional fault diagnosis methods.Manifold learning is a nonlinear dimension reduction method, which can well reveal the intrinsic information of nonlinear data sets, and manifold learning has already been applied to machinery fault diagnosis. According to the characteristics of wind turbine transmission systems, incipient fault diagnosis method for wind turbine transmission system based on generation manifold learning is proposed, at the basis that the theory of manifold learning method is researched and the function of manifold learning is extended. Incipient fault diagnosis method based on generation manifold learning includes nonlinear signal denoising based on unsupervised manifold learning method, weak feature extraction based on supervised manifold learning and small sample pattern identification of incipient transmission system faults, and at last the fault diagnosis of incipient transmission system faults is achieved. The main research contents of the thesis are as follows:(1) Aiming at the problem that the monitoring signals are interfered by strong nonlinear noises, a nonlinear signal denoising method based on unsupervised manifold learning is proposed. The signals are decomposed into several frequency bands by wavelet packet decomposition, and the decomposition coefficients of each frequency band are reconstructed into the high-dimensional phase space. The parameter of phase space reconstruction and the intrinsic dimension of signal are determined according to the signal-to-noise distribution of each signal component. Combined with the advantage of unsupervised manifold learning method in revealing intrinsic information of nonlinear data set, the signal with noise is projected from high dimensional phase space to low dimensional signal space by local tangent space alignment(LTSA), and the method can be applied to nonlinear denoising of signals of transmissions systems.(2) Aiming at the problem that the fault features of wind turbine transmission system incipient faults are weak and hard to extract, weak feature extraction method based on supervised manifold learning method is proposed. Hybrid domain feature set is constructed to comprehensively characterize the early faults. Improved kernel space distance evaluation method is proposed to select the sensitive features and eliminate the interfering and noise features. According to the label information of samples, two full-supervised manifold learning methods are presented, including supervised extended-local tangent space alignment(SE-LTSA) and enhanced supervised locality linear embedding(ESLLE), and a semi-supervised manifold learning method, semi-supervised local tangent space alignment(SS-LTSA), is presented. New sample embedding method is also introduced to embed the new samples into low dimensional feature space for dynamic fault feature extraction. The fault sample label information is organically integrated into dimension reduction process in supervised manifold learning method, which can greatly improve the discrimination of the extracted low dimensional features.(3) Aiming at the problem that incipient fault samples are difficult to collect and lack, small sample incipient fault pattern recognition method based on least square support vector machine(LS-SVM) with optimized parameters is proposed. Enhanced particle swarm optimization(EPSO) is proposed to select the optimal parameters of LS-SVM. By introducing local searching ability, redefining searching speed of particles to involve more information and adaptively adjusting the parameters of optimization algorithm, EPSO can quickly find the global optimal parameters of the fault pattern recognition modal.(4) The system intergration of the proposed generalization manifold learning method is achieved. The system mainly contains the following parts, including data acquisition, signal analysis, early weak fault feature extraction, early fault pattern recognition, and so on. The engineering application of generalization manifold learning method is realized, and the subsystems are tested through examples.At the end of the thesis, the work of this paper is summarized, and expectation of the relative technology development is presented.