Advanced data-driven modelling approaches to alarm-related fault detection and condition monitoring of wind turbines

With the increasing demand of renewable energy, the installation of wind power capacity has been rising exponentially in the past decades. With growing size of modern variable speed wind turbines, the operation and maintenance (O&M) cost can be very high. In order to effectively reduce the O&M cost and maximise the reliability of wind turbines, the condition monitoring has been considered as a most viable solution. However, to improve the reliability of the condition monitoring system, large number of sensors are required and thus numerous data are produced. This can increase the complexity of the monitoring system and bring heavy burden to the computation process. Besides, limited researches has been conducted to study the relationships between alarms and faults. The thesis starts with overview of current condition monitoring technologies and systems. Then the monitoring data used in the research are explained, which include both supervisory control and data acquisition (SCADA) data and data produced from simulation models. A statistical tool based on Kullback-Leibler divergence (KLD) is proposed for feature extraction, considering normal and abnormal behaviour presented in the monitoring data. The proposed method is improved with kernel support vector machine (KSVM), thus capable of classifying the normal, alarm and fault condition of the operational wind turbines. Furthermore, an approach is proposed based on long short-term memory (LSTM) incorporating a KLD for fault detection and identification of representative faults. This method can effectively distinguish the alarms from the faults, from which the distinguished alarms can be considered as an early warning of the fault occurrence. The true positive rate of the proposed LSTM-KLD is 94% whereas the true negative rates for alarm and fault are 96% and 90.9%, respectively. In the end, a wind turbine test rig is designed and developed, from which experimental data are obtained to validate the proposed fault detection algorithms and models. The contributions of the research mainly have three aspects. For fault classification, the kernel function is adapted by both principal component analysis (PCA) and support vector machine (SVM) in order to transfer the linearly inseparable problems into linearly separable problems. Besides, variable selection with kernel PCA is proposed for effective condition monitoring, which can reduce the computation load while retaining the most useful information in the monitoring data. For alarm detection, a statistical tool based on KLD is employed to discover the behaviours of specific components in different operation conditions. By incorporating LSTM with KLD, the fault can be localised by correlating the alarms in the data. With this hybrid method, the fault severity can be estimated based on the alarm signal since it can provide sufficient information as required to indicate early warning of the fault. For experimental validation, a PMSG (permanent magnet synchronous generator) based wind turbine test rig is designed and constructed to emulate the operational behaviours of the turbine under various type of faults in order to collect sufficient experimental data to validate the proposed algorithms for fault detection and severity estimation