168 J. D. Eckels et al. Recent advances in ultrasonic NDE seek to minimize part inspection downtime and improve the accuracy, reliability, and usability of the technique by reducing processing time and utilizing full-field, non-contact, guided wave measurements of the part or component response. Staszewski et al. [3] and Aryan et al. [4] used a 3D laser Doppler vibrometer (LDV) to obtain the surface velocity response of a plate-like structure to ultrasonic excitation from a piezoelectric transducer (PZT). Local cracks or holes in the plate were detected and quantified by alterations in the transient Lamb wave propagation pattern near the location of the defects. Michaels and Michaels identified bonding flaws in an aluminum plate using acoustic wavefield imaging of transient propagating Lamb waves with a non-contact, air-coupled transducer [5]. Studies by Rogge and Leckey [6] and Ruzzene [7] used a 3D Fourier transform of LDV velocity measurements to visualize defects in the frequency/wavenumber domain. Mesnil et al. improved the processing time of these methods by an order of magnitude by performing frequency/wavenumber domain analysis on a steady-state LDV measurement [8]. Acoustic wavenumber spectroscopy (AWS) is a modern ultrasonic NDE technique that overcomes the noise level and delay issues associated with previous transient wave interaction LDV studies by utilizing the full-field, single-tone, steadystate surface response and wavenumber-domain processing to rapidly visualize and characterize defects in a plate-like structure [9, 10]. A single-tone ultrasonic excitation is applied to the surface of a structure by a PZT, and a 2D scanning LDV obtains the steady-state, out-of-plane surface velocity response of the structure. While most LDV measurements rely on transient Lamb wave interactions with defects for damage detection, AWS uses time-invariant properties of the wave, such as wavenumber, to indicate structural damage [9]. The frequency-wavenumber processing described by Flynn et al. [10] produces a full-field map that estimates the local wavenumber on a pixel-by-pixel basis over the scan area of the structure. For a given excitation frequency and wave mode, local changes in wavenumber correspond to changes in plate geometry or material properties, which is often indicative of damage. Plate thickness can be obtained from Lamb wave dispersion relationships, which relate the frequency and thickness to the wavenumber for different wave modes for a given material [11]. Like wavenumber, changes in plate thickness in a local region of a structure can correspond to areas of hidden damage, such as corrosion, indentation, or delamination underneath the surface of the component. Despite its performance improvements from previous ultrasonic NDE techniques, AWS has several limitations. In a recent study that explored the performance limits of AWS, O’Dowd et al. noted that wavenumber estimates lose accuracy as the diameter of a defect decreases below the nominal wavelength of the propagating waves [11]. Similarly, wavenumber estimates are inaccurate along the edges of the LDV scan area. Because the AWS algorithm utilizes a spatial Fourier transform to process data in the wavenumber domain, regions of the scan area with limited wavelength information show limited wavenumber estimation accuracy, including smaller diameter defect regions and the regions around the edges of the scan area. O’Dowd et al. also reported lower accuracy in wavenumber estimation for defects with thicknesses slightly lower than the nominal plate thickness [11]. Regions of similar thicknesses correspond to defects with more subtle changes in wavenumber, which are more difficult to distinguish from surrounding undamaged regions. Obtaining the local thickness of the plate from the local wavenumber and Lamb wave dispersion relationships loses accuracy when there are subtle changes in wavenumber, or when the thickness is closer to the plateau region of the dispersion curves [11]. This method of obtaining the local thickness is less precise and can lead to missed detection of defects with a similar thickness as the nominal plate thickness. Another limitation of AWS is the rate at which computational time scales with increasing scan area. As a result, AWS provides full-field response measurements in larger structures and domains at an increased computational expense. The motivation behind this work was to improve the processing time of AWS and to account for its main limitations in spatial resolution. The limited accuracy of AWS in detecting smaller diameter defects and in regions along the edges of the scan area can be improved by avoiding the resolution artifacts resulting from the spatial Fourier transform. The loss in precision of obtaining plate thickness in regions of subtle changes in wavenumber can be improved by detecting and using plate thickness directly, rather than wavenumber, to indicate and quantify the damage. This work looked into tackling these goals by training a convolutional neural network (CNN) to recognize wave-pattern features directly from a steady-state ultrasonic wavefield image and to classify regions of the image by plate thickness on a pixel-by-pixel basis for damage visualization and quantification, eliminating the intermediate step of wavenumber estimation. This paper will first provide background information on CNNs, image segmentation, and related work. Then, the methodologies used in this study will be explained in detail. Next, several results will be presented, comparing the performance of the CNN model to current AWS technology. Finally, the paper will discuss the major strengths of the CNN model and areas for future work.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==