18 Application of a U-Net Convolutional Neural Network to Ultrasonic Wavefield Measurements for Defect Characterization 173 the plate. In this way, the CNN learned general characteristic features inherent within the different wavefields rather than the specific shape or size of any particular defect. The plate geometry files were created in CAD software and the steady-state ultrasonic excitation response of each plate was obtained using an FEA harmonic response simulation in ANSYS Mechanical. Simulation setup procedures were followed from previous work by O’Dowd et al. [11]. Ultrasonic excitation was simulated by applying a 0.1 MPa pressure at 80 kHz on the top face of the transducer. A constant damping ratio of ζ =0.001 was used in the analysis. The mesh size was tuned to 2 mm to obtain accurate solutions at reasonable computation times. ANSYS ADPL commands were used to export the real and imaginary components of the steady-state harmonic response of the plate for further processing. Finally, ANSYS Python scripting was used to automate the simulation process for all 503 models in the dataset. 18.3.3 Image Processing The real and imaginary components of the steady-state harmonic response of the plate were read directly into MATLAB for initial processing. MATLAB was used to filter out the surface mesh points, interpolate the response to an evenly spaced grid of 400 ×400 mm, normalize the data, and generate a grayscale wavefield image for all of the real component, imaginary component, and magnitude of the steady-state response of the plate for each simulation. Next, MATLAB was used to generate segmentation masks of the plate indicating the location and thickness of defects on a pixel-by-pixel basis from the CAD file geometry. A resolution of 1 mm2 to 1 pixel was used in the segmentation masks. In this study, there were 10 plate thickness classes to identify in each image. Numeric class identifiers from 0 to 9 corresponded to plate thicknesses of 1 mm to 10 mm, with a 10 mm thickness having an identifier of 0. For example, in a given 10 mm thick plate with no defects, every single pixel value in the segmentation label image would have an RGB value of (0, 0, 0), equivalent to its class identifier. If a thickness defect, such as that which would result from corrosion, was introduced to the back of the plate, causing the thickness of the plate in that region to be reduced to 7 mm, then all pixel values in that region of the segmentation label image would have an RGB value of (3, 3, 3). Thus, the CNN model is given wavefield images of a plate and it will be told (via the label images) which wave patterns and regions in the image correspond to which plate thickness value. Figure 18.4 shows an example of a grayscale wavefield image and a segmentation mask for a 10 mm plate with a 1 mm thick hexagonal defect in the upper-right corner. Figure 18.4b is colored for better visualization—10 mm correspond to dark blue and 1 mm correspond to dark red. The hexagonal defect region is evident in the wavefield image where the wavenumber sharply increases compared with the nominal plate wavenumber, indicating a region of the plate with reduced thickness. Only the real-component surface response wavefield images were used in this study. 18.3.4 Data Augmentation The process of generating new data from available data without changing its nature is known as data augmentation [13]. Data augmentation is often used to increase dataset size or is performed on-the-fly between training rounds to improve the accuracy Fig. 18.4 Wavefield and segmentation mask images for a 1 mm thick hexagonal defect. (a) Wavefield image. (b) Ground truth segmentation mask
RkJQdWJsaXNoZXIy MTMzNzEzMQ==