Time-reversal-based ultrasonic non-destructive evaluation using convolutional neural networks

In ultrasonic non-destructive evaluation (NDE) of composite materials or complex-shaped mechanical structures, it is difficult to detect structural defects due to the complex scattering of the excitation signal at the boundary. This study proposes a highly accurate NDE method based on a convolutional neural network (CNN) trained with time-reversal (TR) signals. The approach is applied to honeycomb sandwich panels and skin-stringer structural models, which are widely used in aerospace and space launch vehicles. First, the validity of the finite element model was confirmed by demonstrating consistency between the numerical results and the experimental data obtained from a simple aluminum panel. When the time-reversed, excited signals were applied to the proposed CNN-based classification framework, the model achieved high classification accuracy. Subsequently, the defect classification performance of the proposed algorithm was evaluated using the finite element model for honeycomb sandwich panels and skin-stringer structural models. The robust, high-performance non-destructive testing algorithm was achieved by generating a reconstructed strong signal at the defect candidate location using the time-reversal method, and the appropriate number of receiving sensors and multiple excitation effects were also investigated. As a result, it was confirmed that the high classification accuracy was obtained in all defect areas by acquiring large, reflected signals from defects through time reversal.