Crack characterization from ultrasonic array data via multi-angular-range scattering matrix reconstruction and...

The ultrasonic scattering matrix has demonstrated great potential for characterizing small defects. To enhance crack characterization, this paper proposes a novel framework for scattering matrix denoising, extrapolation, and crack parameter regression. The framework consists of a multi-angular-range scattering-matrix denoising and extrapolation network (MARSM-DENet) and adaptive optimal angular-range scattering-matrix regression network (AOAR-SMRNet). MARSM-DENet performs simultaneous denoising and angular-range extrapolation of scattering matrices measured over a limited angular range, producing low-noise matrices across wider angular ranges. AOAR-SMRNet adaptively emphasizes informative angular range for robust estimation of crack size and orientation. Simulation studies demonstrate that scattering matrices covering the angular ranges of [-75°, 75°] and [-90°, 90°] are most frequently selected as optimal, with average proportions of approximately 62 % and 26 %, respectively, indicating that reconstructions over wider angular ranges can retain more critical defect scattering information and improve noise robustness. Crack size estimation is highly accurate (MAE: 0.043λ-0.048λ, R² > 0.98), while orientation estimation is more sensitive to noise (MAE: 3.34°-7.80°, R²: 0.784-0.971). Experimental validation on twelve crack-like slots with sizes of 0.8λ and 1.2λ and orientation angles within 0°-75° achieves RMSEs of 0.084λ for size and 3.644° for orientation, confirming the practical robustness and effectiveness of the proposed framework in ultrasonic non-destructive testing.