Research On Thermography NDT Of Composites Using Deep Learning Thermal Signal Feature Extraction

Carbon fiber reinforced polymer(CFRP),with its excellent properties of high strength and low density,has been widely used in fields such as aerospace and automotive manufacturing.However,invisible defects such as internal delamination and cracks often occur during its manufacturing and use,which affect the performance and service life of the material.Therefore,it is necessary to detect and evaluate materials by non-destructive testing technology.Optically excited thermography is an efficient and convenient non-destructive testing technique that has been widely used in the detection of CFRP.However,the thermal images acquired by optically excited thermography are susceptible to interference from various factors,resulting in a large amount of noise and affecting defect detection.Therefore,processing algorithms are needed to enhance the extraction of defect information from thermal images and improve the defect detection capability of optically excited thermography.In this paper,the basic theory of thermography defect detection is firstly described,and the theoretical state of the surface temperature field distribution of specimens under long-pulse excitation and lock-in excitation is derived.Through COMSOL finite element simulation software,long pulse thermography and lock-in thermography simulations were carried out on CFRP laminates with flat-bottomed hole defects.Analyze the impact of defect shape on detection and prove that defects with smaller aspect ratios are more difficult to detect.To construct a thermal imaging dataset,an optically excited thermography detection system was built,and long pulse thermography and lock-in thermography experiments were carried out on different CFRP specimens to collect thermal image sequences,which were analyzed using traditional processing methods.This study proposes a model based on a one-dimensional residual attention network(1DRAN)to address the shortcomings of traditional thermal image sequence processing algorithms.This model focuses not on the spatial features of defects but on the differences in temperature evolution between defect and non-defect pixels during the detection process.The results show that the attention mechanism introduced in the model can improve detection performance,and compared to traditional and other deep learning methods,the defect detection performance of this model has also been improved.In addition,the model also predicts the depth of defects,eliminating the need for material physical parameters and addressing the shortcomings of traditional thermal imaging in defect prediction.In order to solve the problem that the number of thermal signals of defective pixels in the dataset is small,by analyzing the disadvantages of Variational Auto-Encoder(VAE)and Generative Adversarial Network(GAN)in generating data,a GAN-VAE fusion model is proposed to realize data augmentation for thermal signal.The effectiveness of the generated thermal signal was determined through visualization and quantitative indicators,and the optimal expansion ratio was determined by using the generated thermal signal as augmentation to improve the detection performance of the 1DRAN model.