Multiscale Wavelet-Driven Transformerfor Blade Icing Detection

Currently, the global demand for green wind power generation is increasing day by day. However, the icing of fan blades can significantly reduce the efficiency of fan power generation, highlighting the increasing importance of efficiently detecting whether the wind turbine is icing. In the traditional approach, the operation status of wind turbines is manually collected. However, with technological advancements, the SCADA system now automatically collects the continuous operational status of wind turbines. Nevertheless, However, the monitored operational data suffers from data imbalance. In this study, we introduce a multi-level neural network built upon the Trans- former architecture. Specifically, to extract multi-level features from the time and frequency domains, we utilize discrete wavelet decomposition. By leveraging the self-attention mechanism of the Transformer, we enhance the feature extraction capabilities for predictive tasks. We investigate data resampling methods to address the data imbalance problem. Comprehensive studies demonstrate that our proposed algorithm improves F1 scores to 90.08%, 81.70%, and 76.21% on datasets processed using data resampling algorithms. The obtained icing detection results validate the applicability of our approach and demonstrate that adopting this algorithm can enhance the accuracy.