Defect detection in infrared thermography by deep learning algorithms
|Abstract:||Non-destructive evaluation (NDE) is a field to identify all types of structural damage in an object of interest without applying any permanent damage and modification. This field has been intensively investigated for many years. The infrared thermography (IR) is one of NDE technology through inspecting, characterize and analyzing defects based on the infrared images (sequences) from the recordation of infrared light emission and reflection to evaluate non-self-heating objects for quality control and safety assurance. In recent years, the deep learning field of artificial intelligence has made remarkable progress in image processing applications. This field has shown its ability to overcome most of the disadvantages in other approaches existing previously in a great number of applications. Whereas due to the insufficient training data, deep learning algorithms still remain unexplored, and only few publications involving the application of it for thermography nondestructive evaluation (TNDE). The intelligent and highly automated deep learning algorithms could be coupled with infrared thermography to identify the defect (damages) in composites, steel, etc. with high confidence and accuracy. Among the topics in the TNDE research field, the supervised and unsupervised machine learning techniques both are the most innovative and challenging tasks for defect detection analysis. In this project, we construct integrated frameworks for processing raw data from infrared thermography using deep learning algorithms and highlight of the methodologies proposed include the following: 1. Automatic defect identification and segmentation by deep learning algorithms in infrared thermography. The pre-trained convolutional neural networks (CNNs) are introduced to capture defect feature in infrared thermal images to implement CNNs based models for the detection of structural defects in samples made of composite materials (fault diagnosis). Several alternatives of deep CNNs for the detection of defects in the Infrared thermography. The comparisons of performance of the automatic defect detection and segmentation in infrared thermography using different deep learning detection methods: (i) instance segmentation (Center-mask; Mask-RCNN); (ii) objective location (Yolo-v3; Faster-RCNN); (iii) semantic segmentation (Unet; Res-unet); 2. Data augmentation technique through synthetic data generation to reduce the cost of high expense associated with the collection of original infrared data in the composites (aircraft components.) to enrich training data for feature learning in TNDE; 3. The generative adversarial network (Deep convolutional GAN and Wasserstein GAN) is introduced to the infrared thermography associated with partial least square thermography (PLST) (PLS-GANs network) for visible feature extraction of defects and enhancement of the visibility of defects to remove noise in Pulsed thermography; 4. Automatic defect depth estimation (Characterization issue) from simulated infrared data using a simplified recurrent neural network: Gate Recurrent Unit (GRU) through the regression supervised learning.|
|Document Type:||Thèse de doctorat|
|Open Access Date:||15 November 2021|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.