Sustainable data-driven framework via transfer learning for icing-detection of high aspect ratio blades

Ice accumulations on structures and rotating blades could introduce significant issues resulting in structure failure, fatigue load increments, safety hazards, etc. Ice detection and anti−/de-icing systems for rotorcraft, aircraft, or wind turbines operating in cold climates become important. This paper introduces a novel ice detection method based on an artificial intelligent technique. The main idea for the proposed ice detection system is that the accumulated ice positions and masses are predicted by the modal property changes, i.e., the natural frequencies. To this end, a deep neural network (DNN) is applied to detect ice mass distributions by considering the variations of the natural frequencies of the slender and flexible rotating and non-rotating blades. To design a refined DNN model, hyperparameter optimization is applied. Furthermore, a transfer learning method is adapted to extend the trained DNN model for the non-rotating blade to the rotating blade. As a result, the parameters related to DNN model are intensively analyzed to design the optimized network. Overall, the proposed method to construct an optimum DNN model as the ice detection system successfully predicts ice mass distributions. In addition, the established DNN model can be easily extended to the new icing scenarios. •A novel and accurate ice detection framework with DNN model for high aspect ratio flexible blades is developed.•The developed ice detection framework is able to predict ice distribution for both standstill and operation conditions.•Optimizing the training data mixing ratio for transfer learning is the one of key elements for transfer learning.