Estimating Damage Detection of an Aircraft Component with Machine Learning Models 135 Limitations The regression models developed demonstrated promising results; however it is important to test their robustness. In order to make sure that our model is able to identify if the structure is damaged and not just matching patterns, more analysis is required. One way to examine this, is to to test the performance of the model with regard to unseen data classes. Therefore two scenarios of our destructive cases were tested, damage cases of 1.6 grams and 4.1 grams removed from the structure were removed from the training data. These tests were done separately, not simultaneously, to see if the model could still predict those damage cases, even though they were left out of the training data. In completing this analysis that is shown in figures 6 and 7, the predictions of damage were not accurate. These results can be interpreted that the model will perform well under the circumstance that it has data from every damage case possible. This also indicates that there is a possibility the model was not identifying the physics of the model and rather trying to pattern match and does poorly when it has not seen similar cases in the training data. Figs. 6&7 SVR INTERPOLATION RESULTS. Conclusion The results demonstrate the application of a damage detection system to an aircraft and practicability of utilizing structural health monitoring for damage detection. However, there are still clear limitations of the models used in these experiments, and still have significant room for improvement. Although a regression model can be fit to experimental testing data with a strong correlation, the model fails to accurately predict data it has not seen before. There are a few ways to mitigate the effects of this flaw in the model. The first is to increase the amount of training data. The experimental damage cases only included 3 different test cases, if increased would likely lead to more robust results. Additionally, if there were more cases given as training data for the model, the less the model must interpolate new data. The second method is to change the damage sensitive features implemented in the machine learning model. The purpose of the damage sensitive features is to extract characteristics of the frequency data that are sensitive to damage. The features used in these regression models are not extracting the necessary information from the frequency data and should be changed to ensure stronger results. Aside from improving the accuracy of this system, another area for improvement would be to expand the capabilities of this system to be able to determine the location of damage alongside severity. Acknowledgments The team would like to acknowledge the support from US Army Program Executive Office - Aviation, PM FARA, and Fulcrum. Specifically, express our gratitude to Dr. Danny Parker, Mr. Kevin McGonigle, Mr. James Hale, Mr. Mike Pollut, Ms. Monna Moser and Mr. Evan Stringer for their advice and support throughout the project. Finally, we would like to acknowledge capstone advisors Dr. Matarazzo and LTC Bellocchio for their unending support for this project and team.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==