Deep learning with filtering for defect characterization in pulsed thermography based non-destructive testing

Abstract

Pulsed thermography is widely used for non-destructive testing of various materials. The temperature profile obtained after pulse heating is used to characterize the underlying defects in an object. In this paper, the automation of the process of defect visualization and depth quantification in pulsed thermography through various deep learning algorithms is reported. Stainless steel plate with artificial defects is considered for analysis. The raw temperature data is smoothed using moving average, Savitzky-Golay and quadratic regression filters to reduce noise. Thermal signal reconstruction, the conventional method to eliminate noise, is also used for generating filtered datasets. Defect visualization refers to identifying and locating the defects in an image sample and Mask region convolutional neural network (Mask R-CNN) is considered for not just detecting the defects but also locating them on the image. The located defects are utilized for depth estimation using the following networks-multi-layer perceptron (MLP), long short-term memory (LSTM) and gated recurrent units (GRU). The input to the networks is the temperature contrast characteristics which symbolizes the difference in temperature over defective and non-defective areas measured over 250 time points and output of the networks is the estimated depth. The study shows that LSTM based approach provides the least percentage error of 5.5% and is a very suitable approach for automation of defect characterization in pulsed thermography.