Determining Airfoil Self-Noise Levels Josh Merchant
Executive Summary The goal of the project is to develop a better way to predict the level of noise that is produced by an airfoil in high velocity air Completed – data preprocessing, model development, general testing and tuning of the model To Do – fine tuning of the model, in depth performance analysis Model is performing as well or better than the target of 0.89 Coefficient of Determination
Data Data was taken from a NASA experiment to determine self-noise levels Data needed to be scaled and centered to allow for learning from all variables Order of the data was randomized Original test run to minimize setup changes Frequency (Hz) Angle of Attack (Degrees) Chord Length (Meters) Free Stream Velocity (m/s) Suction Side Displacement Thickness (m) Sound Pressure (dB) Average 2886.38 6.78 0.14 50.86 0.01 124.84 Sd 3151.52 5.92 0.09 15.57 6.90 Scaled Sd Scaled 1
Method scikit learn MLPRegressor Used CS computers Written in Python Utilized numpy functions Used CS computers Small models, about 1000 samples 10 second run time without graphing Iteratively optimized parameters Made gross changes with scripts to see initial trends Fine tuning the model by hand Current best performance Activation Function – Logistic Nodes – 5 – 15 – 15 – 1 Solver Function – lbfgs Learning Rate – Constant Initial learning Rate – 0.001
Results Model has a Coefficient of Determination of 0.95 Model has a Mean Squared Error of 0.05 Performing better than target network which had coefficient of Determination of 0.89 Mean Squared Error of 0.08 12 or less nodes resulted in underfitting 25 or more nodes resulted in overfitting
Discussion Data processing is Important Tuning Parameters Future Non-Scaled and Non-Centered models had Negative Coefficient of Determination Non-Randomized models had Negative Coefficient of Determination Tuning Parameters Not one dominant parameter Different combinations have similar results Future Tweak Parameters In Depth Analysis