Proceedings of the 2008 International Conference on Electrical Machines Paper ID 844 978-1-4244-1736-0/08/$25.00 ©2008 IEEE Research on a Simple, Cheap but Globally Effective Condition Monitoring Technique for Wind Turbines Wenxian Yang 1 , P.J. Tavner 1 , C.J. Crabtree 1 and Michael Wilkinson 2 1 New and Renewable Energy Group, School of Engineering, Durham University, Durham DH1 3LE, UK E-mail: peter.tavner@durham.ac.uk 2 Garrad Hassan & Partners Ltd, St Vincent Works, Bristol BS2 0QD, UK Abstract- Vibration measurement and lubrication oil analysis are used in wind turbines (WT) as condition monitoring systems (CMS). However, they do not provide a complete solution to the WT CMS problem. The former measurement is sophisticated with high hardware costs, suffering from spurious alarms; the latter monitors the wear and fatigue of gears and bearings, but cannot detect electrical abnormalities occurring in the WT generator and electrical system. So, a simpler, cheaper but moreover globally comprehensive WT CMS is still needed, especially if the WTs are to go offshore, where they are confronted with higher risks and difficulties of access. To meet this requirement, a new WT condition monitoring technique has been researched in this paper. As the WT operates over a widely varying power range, dependant on the stochastic variations of the wind, the monitoring signals are usually non-stationary. In view of this, a wavelet-based adaptive filter is designed to extract the power energy at prescribed, fault-related frequencies which vary with time. The energy information obtained is then used as an indicator of WT condition. The central frequency of the filter is adaptive to the average rotational speed of the generator, and the filter bandwidth depends upon the fluctuation of wind speed. By using this filter, fault features can be extracted whether the WT runs at fixed or variable speed. The proposed technique has been experimentally validated on a WT Test Rig using both synchronous and induction generators as exemplars. Experiments prove that the proposed technique is efficient in assessing the WT condition for both mechanical and electrical abnormalities. I. INTRODUCTION Nowadays, with the extensive use of WTs all over the world, the condition monitoring of these machines is attracting attention from both industry and academia. Currently, almost all condition monitoring techniques being used in the wind industry, e.g. vibration measurement and lubrication oil analysis, are borrowed from other industries [1] where they have achieved success. However, they have not yet proved their success in the wind industry due to the peculiarities of the WT, which has a slow and variable speed. The present vibration-based condition monitoring systems (CMSs) are sophisticated, costly and not necessarily adapted to all types of WTs. The oil particle counter designed for detecting the wear/fatigue of mechanical components is unable to detect electrical failures occurring elsewhere in the WT. However, reliability surveys show that WT electrical system has similar or even higher failure rate than its mechanical system [2, 3]. For these reasons, a simpler, cheaper but moreover globally comprehensive WT CMS is still needed and this paper will report such a technique. From a global view, the mechanical torque and speed of driving shaft and the generator electric power output can respectively be regarded as the input and output of a WT. As a result of the electro-mechanical coupling of the generator, both energy flows will be disturbed by mechanical and electrical abnormalities occurring in WT. Therefore, in a theoretical sense WT condition monitoring and fault diagnosis should be applicable to either of them. However, the torque signal is costly and difficult to extract in practice, so this research will be based entirely on power signal analysis. II. DESIGN OF A WAVELET-BASED ADAPTIVE FILTER The Continuous Wavelet Transform (CWT) has demonstrable merits in processing non-stationary WT signals [5], therefore a wavelet-based adaptive filter, based on the CWT, is designed in this work to extract the power energy at various fault-related frequencies. The CWT of a real-time signal ) (t x is defined as [6] ∫ ∞ ∞ - ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ - = dt a b t t x a a b CWT * ) ( 1 ) , ( ψ (1) where ) (t ψ is the mother wavelet, which is Morlet wavelet in this work. a and b respectively represent the parameters of wavelet scale and time-shift. The asterisk ‘*’ stands for the complex conjugate. Traditionally, the wavelet function ) (t ψ is dilated or compressed continuously by changing the parameter a, so that the signal components within frequency range ω ∈ [0, half sampling frequency] are projected onto appropriate frequencies in the entire time-frequency space. An illustrative example of the CWT is given in Fig.1. It can be easily imagined that this conventional use of the CWT will be computationally intensive. Moreover, most of the calculations are unnecessary for machine condition monitoring, because the fault-related frequencies are few, and the other frequency band does not require analysis. In addition, a computationally inefficient algorithm is less suitable for on- line condition monitoring, than for off-line signal processing. A more efficient algorithm for use on-line would extract only the fault-related frequency components. Two more difficulties limit the application of the CWT in WT condition monitoring. The first is that the conventional CWT manifests the signal in time-scale, rather than in time-