Deep learning for defect characterization in composite laminates inspected by step-heating thermography

This paper presents a complete procedure for the non-destructive analysis of composite laminates, taking advantage of the step-heating infrared thermography and the latest developments of deep neural networks. One-dimensional temperature profiles of the target surface are collected in response to long heat pulses and individually feed a compact network made of convolutional filters, self-tuned to represent the signals in an equivalent feature space of improved discrimination. The resulting features are then classified to obtain the complete three-dimensional characterization of the properties of possible subsurface defects. Experimental validation is proposed to investigate a laminate of glass-fiber-reinforced polymer with several flat-bottom holes by changing the duration of the input heat pulses. This test produces surprisingly good results in the characterization of three classes of defects of increasing depth, including the most challenging at a depth of 6.38 mm, i.e. at the limit of applicability of the step-heating thermography. In the case of an excitation length of 180 s, the average balanced accuracy, precision, and recall are equal to 84.03%, 87.62%, and 82.43%, respectively. Moreover, a threshold operation on the classification scores further boosts the recall values of the class of the deepest defects from 53.87% to 82.41%. This enhancement of sensitivity suggests the applicability of the proposed procedure for the automatic inspection of composites structures in all application fields where safety is mandatory.