Multilayer Input Deep Learning Applied to Ultrasonic Wavefield Measurements

Abstract

Nondestructive evaluation (NDE) methods provide a way to monitor structures which can decrease maintenance/inspection costs, predict damage locations, and indicate the current state of structural health of a system. Acoustic steady-state excitation spatial spectroscopy (ASSESS) is an ultrasonic NDE technique that utilizes surface velocity response data collected by a laser Doppler vibrometer (LDV) to detect and characterize defects. Processing the resultant wavefield images produces a map of damage which is useful in manufacturing and structural health inspection of aerospace/civil structures.

Convolutional neural networks (CNN) were used to estimate plate defects directly from wavefield maps, producing pixel-wise thickness maps which are able to effectively characterize defects without producing boundary artifacts that occur with Fourier-based processing methods. This work utilized a diverse set of geometries and boundary conditions in its simulation-generated training set to ensure generalizability in the CNN. Using the U-net CNN architecture for image segmentation with a pretrained ResNet model as an encoder, a model was trained, validated, and tested which exhibits superior performance in speed and accuracy than traditional methods of defect characterization, while showing versatility over a wide range of plate shapes and boundary conditions.