Influence of lock-in thermography set-up parameters on the capability of a temporal convolutional neural network to characterize defects in a CFRP

Lock-in thermography is a well-established non-destructive technique for detecting damage in composite materials. The success of the lock-in technique in industrial applications depends on several test parameters, such as the excitation frequency, the number of frames per cycle, and the number of excitation cycles that need to be correctly set to reduce testing and processing time. However, quantitative analysis to characterize defects using lock-in thermography is still a challenging task. Machine and deep learning algorithms can be useful to automatize the classification of defects in terms of size and depth. In this regard, the aim of this work is to investigate the influence of lock-in thermography set-up parameters on the capability of a temporal convolutional neural network to characterize defects in a carbon fiber-reinforced polymer specimen. The performance of the proposed neural network in identifying and classifying the depth of defects was evaluated as a function of the number of cycles, the frames per cycle, and the random frames lost during the acquisition of the thermal sequence by an infrared camera. The achieved results have been critically discussed through qualitative and quantitative analyses.