Application of Adaptive Learning Networks for the Characterization of Two-Dimensional and Three-Dimensional Defects in Solids

The objective of the work was to develop an ultrasonic inversion procedure which (1) discriminates, (2) sizes,. a.nd (3) determines the orientation of two-dimensional (crack-like') and· three-d'imensional (void-like) defects in materials. Adaptive learning networks (ALN's) were used to. esttmate dfrectly the defect size and orientation parameters from the spectrum of the echo transient. A 19-element hexagonal synthetic arr·ay measured the scattered fie 1 d within a 60-degree solid ang.l e aperture. The ALN' s were trained on theoretically· generated spectral data where the crack forward scattering model was based on the Geometrical Diffra.ctfon Theory and the void model was· based on the exact Scattering: Matrix. Theory. The theoretica·l'ly trained models were evaluated on both theoretical and1 experimental data. Excellent results were obtained, and the errors for size and odentation estimates were, in general, less than 10%. The significance of tlris work is that~ (1) the ALN approach to: defect characteristics provides a systematic procedure for discovering re·lationships in the data• which could otherwise be overlooked, and (2) signific.ant economic benefits. can be gained by simulating difficult-to-produce defect reflector scenarios. Furthermore, a result of this work has been the development of an a 1 gorithm which can ul t imately be applied fn field and· industrial use. SUMMARY OF RESULTS, CONCLUSIONS, AND RECOMMENDATIONS Table 1 pres·ents the relative errors for the nine. ALN· models developed in this study and evaluated on both theoretical and experimental data. Model's. 1 thr:-ough 4 were trained to estimate the four parameters of the 3-dimensional spheroid models; 5 through B were trained to estimate the four parameters of the 2-dimensional crack; Model 9 was trained to discriminate between the spheroid and crack and, thus, to act as a selector for the appropriate size model. Table 1. ALN Model Results Summary Parameter I Theoretical ! E~~~~~al! Model Model Relative No. Typ