Wind Turbine Fault Diagnosis Based on Cyclostationary Feature Detection From Compressive Samples

Compressive sensing theory, capturing fewer data points without compromising essential information, holds significant promise for high-sampling-rate applications, particularly in wind turbine monitoring. The primary strategy within current compressive sensing fault diagnosis involves designing a dictionary closely aligned with fault characteristics and utilizing atomic decomposition to match feature waveforms. However, these methods exhibit diminished performance in scenarios where fault features are weak, and noise interference is pronounced. We exploit a novel fault diagnosis method within the compressive sensing framework to address this challenge. In this approach, noise-robust statistics, specifically the cyclic spectrum, are estimated using the modified strip spectral correlation analysis (SSCA) algorithm. The crucial demodulation step in the SSCA algorithm entails spectral reconstruction by solving a sparse representation optimization problem. The computational complexity of this method is comparable to the original algorithm, while retaining its noise immunity performance. Through numerical and engineering data analysis, it is demonstrated that similar feature detection results can be attained in compressed samples compared with conventional Nyquist sampling.