Convolution Neural Networks Based Automatic Subsurface Anomaly Detection and Characterization in Quadratic Frequency Modulated Thermal Wave Imaging

Recent trends in thermal non-destructive testing focusing on artificial intelligence and various deep learning architectures have been investigated for quality assessment of different materials. The present work introduces three famous computer vision models (AlexNet, GoogleNet and VggNet) with one-dimensional convolution layers for defect detection for material inspected by quadratic frequency modulated thermal wave imaging. These models employ sequential convolution operations and pooling on temporal thermal profiles and extract deep features further to classify defect and sound regions in the test sample. The three deep learning models are trained from scratch with the experimental thermographic data of a carbon fiber reinforced polymer (CFRP) specimen with artificially simulated flat bottom hole defects of different sizes at varying depths. The performance metrics conclude that AlexNet presents high testing accuracy and F-score of 98.92% and 0.954 resulting in less deviation to the actual labels favoring enhanced defect signal-to-noise ratio with less computation time in CPU-based hardware. Further, the depth of the detected defect was quantified using a recently introduced quantification model using the chirp-z transform-based phase analysis. The estimated depths are rearranged in the respective locations and visualized the depth map.