A Review of Artificial Intelligence Applications in Thermographic Non-Destructive Testing

Thermographic Non-destructive Testing (TNDT) has gained increasing popularity in various industry fields. It can provide rapid, non-contact, and robust non-invasive detection of both surface and subsurface damage. Artificial Intelligence (AI) is an emerging subject that shows increasing potential in almost all fields and has recently attracted significant interest in TNDT. Thermal signals from TNDT have relatively low signal-noise-ratio (SNR), and most thermal images have the common weakness of edge blurring. The abovementioned obstacles lead to high requirements of field expertise and subjectivity in TNDT inspections. One of the purposes of developing AI is substituting human work more efficiently and objectively. The abovementioned weaknesses in TNDT may seek a way out of AI technologies. This paper offers a review of state-of-art researches on AI deployment in TNDT, discussing the current challenges and a roadmap for application expansion. Deep Learning is the most commonly used AI technology since it has powerful feature extraction and pattern recognition capabilities for imaging processing and computer vision. Most existing research adopted Convolutional Neural Network (CNN) models utilizing only spatial information in thermal images to detect defects such as U-net, VGG, Yolo, etc. Except for defect detection, automated defect depth estimation is another focus in the deep learning method. Recurrent Neural Networks (RNNs) such as LSTM and GRUs are usually applied for extracting the temporal feature from thermal sequences, which is sensitive to defect depth. Furtherly, different deep model variations and integrated algorithms are also reviewed, which improves the performance of defect detectability. Another exciting aspect of learning models using 3-dimensional thermograms is their ability to consider the spatial and temporal features to reveal more hidden defects. Some other points, such as the training dataset, which plays a crucial role in making a robust deep model, are discussed at the end of this paper.