Research On Interpretable Deep Learning On Aerodynamic Data

The application of machine learning and deep learning in the field of aerodynamics has gradually emerged in recent years.Researchers have been trying to explore the combination of deep learning and aerodynamic data modeling,and have achieved some meaningful results,but there are still many difficulties that remain unsolved.Due to the lack of interpretability,aerodynamic models based on deep learning are often regarded as black-box models,and no one can know exactly what the model makes decisions based on and whether the decision is reliable.Although the model is mathematically clear and transparent,the physical meaning of the model is often unclear when applied to a specific scene,resulting in the model not being recognized by experts in aerodynamics.In order to improve the interpretability of deep learning models for aerodynamic data,establish a trust relationship between ordinary users and decision-making models,and eliminate the potential threats of aerodynamic data models in practical deployment applications,this thesis first studies the relevant theoretical and technical foundations of machine learning interpretability,especially focusing on the Local Interpretable Modelagnostic Explanations(LIME),and then applies it in deep learning-based modeling of aerodynamic data to explain the black-box prediction model.On the basis of the original interpretation model,this thesis proposes three improved models for LIME---Uniform LIME,GANLIME and CGANLIME,to make up for the problem of ununiform distribution of disturbed samples,the problem of ignoring feature correlations in the process of generating disturbed samples,and the problem of uncontrollable generation of disturbed samples when original LIME solves regression problems with tabular data.For the M6 transonic airfoil data set,this thesis uses the original interpretation model and three improved models to interpret the results of the deep learning-based black-box aerodynamic model for predicting the lift coefficient,drag coefficient and torque coefficient,and compare the local interpretation results and the global interpretation results to qualitatively understand the pros and cons of the model.And quantitatively evaluate different interpretation models by calculating the model evaluation metric,and compare the quality of the models in detail.The experimental results show that LIME can effectively explain the aerodynamic data model based on deep learning.The three improved models make up for the defects of the original LIME,and obtain better interpretation results and higher evaluation metric.