Automated Ultrasound Data Processing for Defect Detection and Characterization Through Machine Learning

本项目将通过两种方法探索CFRP的自动PAUT数据解释的实施：I）开发低延迟深度神经网络（DNN）以在执行扫描的同时动态分析A扫描数据，用于几何特征识别、时间序列的自动选通和缺陷检测，以及II）开发用于图像分析的多任务网络（MN）[2]，在每个B扫描、D扫描和C扫描投影上检测几何特征/缺陷，并在组合阶段交叉验证发现。应用于A扫描数据的实时DNN将用于为检查期间标记的缺陷提供警告，而MN通过由数据的多视图分析授权的不同相关任务进行潜在改进的学习，应该能够以更高的置信度检测缺陷。

This project will explore the implementation of automated PAUT data interpretation for CFRPs through two approaches: I) developing a low latency Deep Neural Network (DNN) to analyze the A-scan data on the fly, while the scan is being performed, for geometrical feature recognition, automated gating of time series, and defect detection, and II) developing a Multitask Network (MN) [2] for image analysis, detection of geometrical features/defects on each B-scan, D-scan, and C-scan projection, and cross-validation of findings at the combination stage. The real-time DNN applied to the A-scan data will serve to provide warnings for defects flagged during the inspection while the MN, with a potentially improved learning through different related tasks empowered by the multi-view analysis of the data, should be able to detect the defects with higher confidence.