The objective of this dissertation research is to investigate both model-based and data-driven fault diagnosis and prognosis approaches for wind turbine hydraulic pitch systems. For modern wind turbines, operation and maintenance (O&M) cost has contributed a major share in the cost of energy (COE) for wind power generation. Condition monitoring can help reduce the O&M cost of wind turbine. Among all the wind turbine components, the hydraulic pitching system is considered in this study for the fault diagnosis and prognosis. Hydraulic pitching system is critical for energy capture, load reduction and aerodynamic braking. Its reliability and maintenance is thus of high priority. The faults of cylinder leakage, air contamination, and valve blockage and silting in hydraulic pitching system are studied in this dissertation research.;Detection of the aforementioned faults is first studied in a model based approach based upon the nonlinear dynamic model of the hydraulic pitch system. The valve blockage is detected by estimating the effective valve orifice area using the valve actuation model. Air contamination is detected by observing the change of bulk modulus for the system. Based on the nonlinear dynamic model of the hydraulic pitch system, the recursive least-square and adaptive parameter estimation algorithms have been developed to identify the effective bulk modulus, internal and external leakage coefficients. The convergence of the adaptive algorithm was proved with Lyapunov analysis. These schemes can, not only detect, but also isolate individual faults from each other in spite of their coupled relationship in the hydraulic model, which is advantageous for practice. In parallel, a data-driven approach is also investigated. The fault and not-fault conditions for the internal leakage in the hydraulic system are classified through the self-learning asymmetric support vector machine (ASVM) algorithm, which can maintain the complexity of the fault model. The improved ASVM algorithm can adaptively select the minimal number of support vectors while maintaining the desired classification performance, which makes the practical implementation of the classifier computationally more efficient. The prognosis, i.e. prediction of the remaining useful life (RUL), has been studied for the valve silting with the modified hidden semi-Markov model (HSMM). Improvement has been made to the existing HSMM method which can well handle the underflow issue with relatively large number of observation samples.;The proposed methods are first verified through the simulation study based on the aerodynamic loading on the pitching axis under smooth and turbulent wind profiles obtained from the simulation of a 1.5 MW variable-speed turbine model on the FAST (Fatigue, Aerodynamics, Structural and Tower) software developed by the National Renewable Energy Laboratory (NREL).;A scale-down experimental setup has been developed as the hydraulic pitch emulator, with which the proposed algorithms can be verified through experimental data. The setup consists of two back-to-back hydraulic cylinders, with one emulating the pitch cylinder and the other emulating the pitching-axis load. The pitching-axis load inputs are obtained from simulating a 1.5 MW variable-speed-variable-pitch turbine model under turbulent wind profiles on the FAST. Limited by the available experimental resources, only the leakage faults can be realized on the setup. With the same learning rate and small external leakage in the simulation, the estimation errors increase when the external leakage increases from 0.05 to 0.16 liter/min. When the internal leakage increases from 0.5 to 0.8 liter/min, the estimation errors also increase. When both simulated internal and external leakage are larger than 1.5 liter/min in the experiments, the mean estimation errors can be less than 12% with varying learning rate in each case. With the experimental data, the leakage and leakage coefficients can be predicted via the proposed method with good performance.;With the experimental data, the developed self-learning ASVM algorithms have also been applied to the detection of internal leakage fault from the normal conditions, with all the experimental data correctly classified.
展开▼
