Machine Learning for US Army UAVs Sustainment: Assessing Effect of Sensor Frequency and Placement on Damage Information in the Ultrasound Signals

US Army unmanned aerial vehicles (UAVs) in the future will be sustained for longer durations if damage in structural parts is continuously monitored from the damage-inception stage and continuously through vehicle life. Neural networks based machine learning (ML) are proposed, demonstrating that the length of a developing fatigue crack can be estimated continuously using the ultrasound signals. Using a 0.5-TB data set that was obtained from a carefully selected set of experiments, the ML was developed in three stages: 1) feature development, 2) outlier elimination and 3) role of the excitation frequency and exciter-receiver path in the ML of the crack length. In the first stage, the recorded 8000-point ultrasound signals were reduced, first, to 63 features comprising the major statistical features of the returned signal and the seven scales of a seven scale wavelet decomposition of the returned signal. Using an autoencoder algorithm, outliers in the input were identified and removed. A four-layer, 63-32-16-1 neural network based linear regression algorithm was used to predict the crack length from the input features. The results indicated that the ML algorithm gave correlation in the range of 99.43-99.97% when both the exciter-frequency and the exciter-receiver paths are fixed. For investigating the effects of the excitation frequency and the exciter-receiver path on the crack-length information in the returned signal, a similar neural network algorithm was used. One or two additional variables were added to the incoming samples' feature space depending on whether the excitation frequency or the exciter-receiver path or both were variables. ML for crack-length estimation showed promise for these situations, too. In the more practical first situation, where the exciter frequency is fixed and the exciter-receiver path is uncertain, the algorithm showed an accuracy in the range of 96.97-98.92%. This algorithm still gave a correlation above 85% when there was uncertainty in the excitation frequency and exciter-receiver paths, as well. This work thus demonstrates the potential for monitoring fatigue crack length growth throughout the life of a vehicle for an increased sustainment of the US Army UAVs.