Blade ice detection encounters challenges such as imbalanced data distribution and data overlap in actual operation. These lead to the poor performance in the accuracy of ice detection. To deal with such issues, this paper introduces a hybrid model that combines a bidirectional long short-term memory network (BiLSTM) with a fully connected neural network (FCNN). First, a feature reconstruction technique that integrates physical mechanisms with data-driven methods is utilized to generate new features. They are combined with icing-sensitive features to create an enhanced feature dataset. Then, the BiLSTM is employed to extract temporal features from the dataset, which are further analyzed by the FCNN to enhance feature extraction. To address challenges related to data distribution overlap and class imbalance, the proposed model incorporates both center loss and focal loss functions. Experimental results show that the proposed approach achieves good performance in detecting blade ice and could identify both the presence and severity of ice on wind turbine blades effectively.

wind turbine, blade ice detection, feature reconstruction, data-driven