Interference reduction of high-energy noise for modal parameter identification of offshore wind turbines based on iterative signal extraction

An iterative noise extraction and elimination method is proposed, to solve the difficulty of modal parameter identification caused by contaminated high-energy components in measured signals. The approach is based on the idea that if only the high-energy noise is extracted and eliminated accurately, modal parameter identification can be improved significantly by using the remaining (filtered) signal. A theoretical development is that a general form of noises is considered, i.e., they can be purely harmonic or damped, which implies noises from machine vibration or other unknown sources can be taken into account. The advantage of this approach is that only one significant, relatively noisy component is extracted at each iteration, i.e., only the most reliable component is extracted each time, so more accurate high-energy noise elimination is expected. To demonstrate the accuracy of the proposed method, a numerical signal with one high-energy noisy component is used. Numerical results show that the approach can extract and eliminate high-energy noise at 2.1 Hz accurately by setting a small model order. A special case is the noisy component as a harmonic, which means the traditional Fourier transform can be applied. However, one can conclude that the proposed method outperforms the Fourier transform due to its limitation of fixed frequency resolution. To further investigate the effectiveness of the approach, an experiment on a steel offshore platform is conducted, and experimental results indicate that when measured raw data is used, only the first mode can be identified; if filtered data is used by employing the proposed method, the first two modes can be identified with improved damping ratio estimation. Finally, sea-test data from an offshore wind turbine in operating conditions are used, results show that the first mode with a frequency of 0.3592 Hz and a damping ratio of 0.0151 are successfully identified, while no reasonable parameters can be obtained by employing the stochastic subspace identification method, even when the used model order reaches 300.