Defect shape detection and defect reconstruction in active thermography by means of two-dimensional convolutional neural network as well as spatiotemporal convolutional LSTM network. Issue 2 (15th March 2022)

• Record Type: Journal Article
• Authors: Müller, David; Netzelmann, Udo; Valeske, Bernd
• Is Part Of: Quantitative infrared thermography. Volume 19:Issue 2(2022)
• Page Start: 126
• Page End: 144
• Publication Date: 2022-03-15
• Subjects: Defect detection -- defect profile reconstruction -- neural networks -- induction thermography -- active thermography; Infrared technology -- Periodicals; Thermography -- Periodicals; Infrared technology; Thermography; Periodicals; 621.362
• DOI: 10.1080/17686733.2020.1810883
• Languages: English
• ISSNs: 1768-6733

Abstract

ABSTRACT: A neural network (NN) for semantic segmentation (U-Net) was used for the detection of crack-type defects from thermography sequences. For this task, data sequences of forged steel parts were acquired through induction thermography and the corresponding phase images calculated. The results for defect detection were quantitatively evaluated using Intersection over Union (IoU) metric. Further, a combination of 2D convolutional layer as well as LSTM (Long-Short-Term-Memory) is shown, which includes three-dimensional aspects in the form of time dependent and spatial changes and allows a defect shape reconstruction of back wall drillings. Therefore, pulsed thermography sequences were simulated with COMSOL Multiphysics. Finally, the reconstruction results were compared with the ground-truth defect profile using Mean Squared Error (MSE). The approaches provide improvements over conventional methods in non-destructive testing using infrared thermography.