Nondestructive detection and analysis based on data enhanced thermography

Abstract

Abstract Machine learning and data fusion methods have played a significant role in the nondestructive testing (NDT) of composite materials in modern industries. Among them, feature extraction thermography models such as principal component thermography have been well studied and applied due to their advantages of dimensionality reduction and discriminative representation. However, current pulse thermal imaging technology can only collect limited images since the tested materials cannot usually be overheated, which also decreases the performance of the feature extraction thermography models. In this paper, an improved denoising convolutional autoencoder with U-net architecture is proposed for data enhancement purposes. Using the developed structure, enhanced images are generated, in which the deep spatial information in the original images is maintained while measurement noises are decreased simultaneously. Using the original and enhanced images together, the NDT performance based on feature extraction thermography is further improved. Finally, the feasibility and effectiveness of the proposed methods are demonstrated by two defect detection experiments on carbon fiber reinforced polymer.