Research On Methods Of Condition Diagnosis For Wind Turbine Drive Train

The loss of generated energy and maintenance expense brought by wind turbine failures has become the main part of wind farm operation cost. The relevant faults of wind turbine drive train (WTDT) are the leading causes to wind turbine downtime and its maintenance cost is very high. Research on methods of condition diagnosis for WTDT is of great significance for wind farms to make proper maintenance strategies, which can increase wind turbine availability, reduce wind farm operation cost and strengthen the competitiveness of wind power in the field of renewable energy.With WTDT as research object, based on wind turbine SCADA data and vibration data from key components, this dissertation mainly discussed the methods of condition monitoring and failure diagnosis for WTDT, aiming for the real-time monitoring of operaton condition, the failure pre-diagnosis, as well as the fault location and performance degradation assessment. The main contribution of the dissertation can be summarized as follows:(1) A multi-parameter alarm mode for WTDT was proposed. Taking various operation parameters comprehensively into consideration, the K-means clustering algorithm was used to classify wind turbine operation conditions; the Euclidean distance between the center of the cluster and each sample was calculated as the new alarm threshold in each working condition subspace. Compared with conventional single-parameter mode, the multi-parameter alarm mode for WTDT had considered various operation parameters from SCADA system in both "condition classification" and "threshold setting". Experiments showed that the proposed method could reduce the turbine false alarm rate effectively.(2) A WTDT condition monitoring model based on least squares support vector machine optimized by gravitational search algorithm (GSA-LSSVM) was established. The active power, wind speed, nacelle temperature, rotation speed of main shaft and gearbox oil temperature (GOT) from SCADA system when wind turbine is in healthy operation condition are selected to establish the standard expert database; the combination of penalty coefficient and kernel parameter of least squares support vector machine (LSSVM) were optimized by gravitational search algorithm (GSA), and the GOT mapping model was established based on GSA-LSSVM model. The definition of distinguish index is the ratio of the predicted GOT values over the measured GOT values, and by means of monitoring the statistical characteristics of distinguish index, the condition monitoring of WTDT could be realized. The reliability and practicability of the proposed model was verified after experiments.(3) A WTDT failure diagnosis method based on artificial neural network optimized by gravitational search algorithm (GSA-ANN) was proposed. The gravitational search algorithm (GSA) was applied to optimize the initial weight and threshold values of artificial neural network (ANN) in order to improve the stability of the network. Due to the non-stationarity and complexity of vibration signals from WTDT, the power spectrum entropy, wavelet entropy, box dimension and correlation dimension, skewness and kurtosis were extracted as features of the fault diagnosis. Based on the GSA-ANN model, with features such as power spectrum entropy as inputs, the pattern recognition of WTDT failures was conducted, and the results were compared with that of artificial neural network optimized by particle swarm optimization algorithm (PSO-ANN). The calculation results showed that the GSA-ANN method could accurately identify the gear wear, gear tooth breaking faults in gearbox and generator bearing looseness, which were three typical WTDT faults, with superior diagnosis performance and calculation efficiency than PSO-ANN method.(4) A WTDT failure diagnosis method based on least squares support vector machine optimized by glowworm swarm optimization algorithm (GSO-LSSVM) was proposed. The failure diagnosis model GSO-LSSVM was established with the optimization of the penalty coefficient and kernel parameter of least squares support vector machine (LSSVM) by glowworm swarm optimization algorithm (GSO). The power spectrum entropy, wavelet entropy, box dimension, correlation dimension, skewness and kurtosis were extracted from rolling bearing vibration signals as the features of the fault diagnosis. Based on GSO-LSSVM model, the pattern recognition of rolling bearing failures with different locations and degradation degrees was conducted, and the results were compared with that of least squares support vector machine optimized by genetic algorithm (GA-LSSVM). The calculation results showed that, GSO-LSSVM could effectively locate the rolling bearing faults and accurately assess the degradation degree of bearing performance, at the same time superior to GA-LSSVM method in both the diagnosis accuracy and computation efficiency.(5) A real-time health condition evaluation model for WTDT model was established based on gaussian mixture Copula model (GMCM). The Copula function and gaussian mixture model (GMM) were combined, and with mean values and kurtosis from rolling bearing vibration signals as features, based on GMCM the negative log likelihood probability could be calculated and the degradation index could be obtained, which realized the real-time health condition evaluation of rolling bearings. Experiments showed that GMCM was superior to GMM in the pre-detection of rolling bearing degradation trend, which could help with the early forecast of potential failures and to prevent fault deterioration. Finally the research achievements of the condition diagnosis for WTDT were integrated and applied in the "System of On-line Operation Condition and Performance Assessment for Large-scale Wind Turbines".