Implementing the conventional total focus method(TFM)for visualizing internal damage in reinforced con-crete(RC)is beset with computational challenges and a high dependence on physical principles.To overcome these cha...Implementing the conventional total focus method(TFM)for visualizing internal damage in reinforced con-crete(RC)is beset with computational challenges and a high dependence on physical principles.To overcome these challenges,an efficient total focus imaging method based on deep learning is proposed.This method deals with array ultrasonic time-domain signals from cracked RC beams.A deep neural network(DNN)employing a feature extraction+multilevel feature fusion+matrix construction architec-ture was developed;this architecture enabled the DNN to learn the underlying physical principles of the TFM.The ar-chitecture effectively transformed ultrasonic time-domain signals into a B-scan matrix.Training,validation,and test-ing data were collected by measuring eight RC beams with preset artificial cracks using a low-frequency shear wave ar-ray ultrasonic system.The results demonstrated that the re-constructed B-scan matrices had a peak signal-to-noise ratio of 26.94 dB and a structural similarity index of 0.978.Fur-thermore,the proposed method required 42%fewer floating-point operations compared with physics-based cal-culations,achieving total focus imaging with lower compu-tational cost.The study facilitates the advancement of ultra-sonic total focus imaging of RC structures from physics-based methods to data-driven methods without re-quiring prior physical knowledge,thereby providing robust support for further nondestructive evaluation and quantita-tive analysis.展开更多
基金Science & Technology Specific Project of Jiangsu Province (No. BZ2024047)Key R&D Program of Ningbo (No. 2024H013)the National Natural Science Foundation of China (No. W2412092)。
文摘Implementing the conventional total focus method(TFM)for visualizing internal damage in reinforced con-crete(RC)is beset with computational challenges and a high dependence on physical principles.To overcome these challenges,an efficient total focus imaging method based on deep learning is proposed.This method deals with array ultrasonic time-domain signals from cracked RC beams.A deep neural network(DNN)employing a feature extraction+multilevel feature fusion+matrix construction architec-ture was developed;this architecture enabled the DNN to learn the underlying physical principles of the TFM.The ar-chitecture effectively transformed ultrasonic time-domain signals into a B-scan matrix.Training,validation,and test-ing data were collected by measuring eight RC beams with preset artificial cracks using a low-frequency shear wave ar-ray ultrasonic system.The results demonstrated that the re-constructed B-scan matrices had a peak signal-to-noise ratio of 26.94 dB and a structural similarity index of 0.978.Fur-thermore,the proposed method required 42%fewer floating-point operations compared with physics-based cal-culations,achieving total focus imaging with lower compu-tational cost.The study facilitates the advancement of ultra-sonic total focus imaging of RC structures from physics-based methods to data-driven methods without re-quiring prior physical knowledge,thereby providing robust support for further nondestructive evaluation and quantita-tive analysis.
