A 3D reconstruction method of defects inside flat ceramic membranes is proposed, basing on dynamic array element synthesis aperture focusing ultrasound imaging, in order to address the problems of poor visualization of defects inside flat ceramic membranes and to reduce the data requirement for ultrasonic 3D reconstruction. In the dynamic array synthetic aperture focused ultrasound imaging method, a multi-array synthetic aperture in defective areas and a single-array synthetic aperture in non-defective areas are used. By using this method to acquire B-scan images of defects in flat ceramic membranes at multiple location sequences and a body drawing method, three-dimensional visualization of defects in flat ceramic membranes is achieved. It is found that relative errors of the reconstruction of scratches using the dynamic array element synthetic aperture focused ultrasound imaging method range from 1.27% to 2.7%, while the relative errors of the reconstruction of holes range from 2.38% to 3.03%, with an average increase in reconstruction speed of 26.57%. The three-dimensional reconstruction method makes the defects more intuitive and provides an objectiv e condition for subsequent defect analysis. The method is well adapted and economical, with potential applications in the field of non-destructive testing.

For the flowing hole affecting the one-dimensional convolutional neural network to identify the ultrasonic defect signal inside the flat ceramic membrane, this study proposed a 1D-CNN based on error compensation for the ultrasonic defect signal identification method of flat ceramic membrane. First, the ultrasonic flaw detector was used to scan the flat ceramic film and obtain the ultrasonic signal of the flat ceramic film. Second, through the analysis of the pulse reflection method, the inherent position of the flowing hole that causes the flowing hole ultrasonic signal was generated with the movement of the probe and change. A "rectangular box" was used for the ultrasonic signal error compensation of the generated flowing hole. Finally, the error-compensated ultrasonic signals were learned and classified employing a 1D-CNN model involving a fused attention mechanism. The experimental results demonstrated that the accuracy of the proposed 1D-CNN based on error compensation for ultrasonic defect identification of flat ceramic films was 95.63%, which was 17.06% higher than that of the 1D-CNN model without error compensation. Thus, the proposed detection method indicates promising potential and value in industrial applications.

针对平板陶瓷膜表面缺陷实时检测时存在检测准确率较低的问题,本文提出了一种融合坐标注意力和自适应特征的YOLOv5陶瓷膜缺陷检测方法.通过在原有YOLOv5模型的主干网络中加入坐标注意力机制,建立位置信息和通道之间的关系,从而更准确地获取感兴趣区域.在原始网络的预测网络中融入自适应特征融合机制,提高模型对多尺度缺陷的检测能力.将空洞空间卷积池化金字塔模块替换原始网络中的空间金字塔池化模块,提高卷积核视野获取更多的有用信息.实验结果表明:本文模型平均精度为97.8％,检测帧数为32 FPS,平均精度与原始YOLOv5模型相比提高了5.5％.本文提出的模型在满足平板陶瓷膜缺陷的实时检测条件下,提高了模型的检测准确率,对推动平板陶瓷膜缺陷检测的发展具有一定的参考价值.