A Neural Network Algorithm for Ultrasonic Imaging of Pile Defects

Ultrasonic testing has been widely used in the quality inspection of pile foundations due to its non-destructiveness and high accuracy. Traditional ultrasonic transmission method relies on two-dimensional waveform diagrams （e.g.， sound velocity， amplitude， and PSD） for defect judgment， and the readability and intuitiveness of the detection results are relatively poor. Based on a modified residual convolutional neural network algorithm （M-ResNet）， an original ultrasonic dataset using the ray-tracing method was constructed. Deep learning technology was employed to train the network model， and MATLAB was used to develop code for displaying 3D pile foundation defect images and detection results. Evaluation indicators （ICC and CCC） for defect images and algorithm noise immunity tests demonstrated that the algorithm operated rapidly with high detection accuracy， could directly display the defect range and size， and showed more advantages in imaging compared to traditional methods. The proposed method can directly use existing instruments and test data from conventional ultrasonic methods to achieve 3D visualization and defect detection， with research results applicable to similar engineering inspections.