Prediction of Glaze Ice Horns from Aerodynamic Degradation Through Logistic Regression

In-flight, real-time ice detection enhances the pilot’s airframe icing awareness during aircraft and unmanned aerial vehicle operations in adverse weather. The objective of this work is to develop logistic regression–based binary and multiclass classifiers for ice class predictions to determine the onset of glaze ice horns on an airfoil using the degradation of aerodynamic performance during icing at mid- and low-Reynolds-number conditions. Multiclass classifiers divided the glaze ice horns class into incipient and developed. Cross-validation results yielded validation accuracies of 97% and 71% for ice horns onset detection accuracy. The high validation accuracy indicates that the instantaneous ice class can be inferred from the instantaneous aerodynamic performance. The glaze ice horns onset accuracy of the models was improved by 9% by overfitting the streamlined rime to glaze ice horns transition and by using the aerodynamic performance time evolution as input. It was observed that a larger training batch reduced the glaze ice horns onset accuracy when using overfitting strategies. The trained models were used in virtual test cases showing that the improved multiclass classifier can anticipate the glaze ice horns onset by sensing a variance increment of the glaze ice horns probability.