Research On Infrared Thermal Wave Detection And Convolutional Neural Network Recognition Of HSCs Defects

Honeycomb Sandwich Composites(HSCs)are widely used in primary and secondary structural components of aerospace vehicles because of their high specific stiffness,high specific strength and low density.The process of preparing HSCs is complicated,and during the preparation process or long-term operation in harsh environment,the debonding defects between skin and adhesive layer,adhesive layer and honeycomb core may occur,which seriously affects the normal use of related components.Therefore,the fast and accurate detection and identification of internal defects in HSCs has become a hot and difficult research area.Deep learning based convolutional neural network has the advantages of strong feature extraction ability,good robustness and high recognition efficiency,which is widely used in image processing,image classification,target recognition and other fields.In this paper,deep learning is combined with traditional infrared nondestructive testing technology,and a neural network model based on lock-in thermography algorithm for two-channel imaging is proposed to conduct an in-depth study on the identification,classification and quantitative detection technology of HSCs defects.Firstly,a three-dimensional heat conduction model of a multilayer structure with a sinusoidal regular form of heat flow acting on simulated HSCs is established,and the amplitude and phase of the temperature signal generated by the modulated heat flow on the heated surface are solved.The feature images of the IR image sequences acquired by the thermal imaging camera are extracted using the lock-in thermography algorithm,and the feature images are processed using contrast enhancement,filtering processing,image segmentation,and data enhancement,etc.The effects of different size filter windows on the signal-to-noise ratio are explored,and the applicability of different image processing methods is analyzed.Secondly,the traditional lock-in thermography algorithm is combined with convolutional neural network to build a convolutional neural network model for HSCs defect recognition based on deep learning,and a more efficient network structure for HSCs defect recognition is explored.Using the pre-processed HSCs feature image dataset for training,the influence and pattern of neural network parameters on the defect recognition accuracy are studied and elucidated by experimentally modulating the neural network parameters.The HSCs defect recognition software was developed and programmed using Python language to achieve fast,efficient and accurate recognition of HSCs defects.Subsequently,the influence of different improvement methods on the accuracy of defect recognition under complex environment and the recognition speed is studied through experiments.The data sets of several types of typical debonding defects of HSCs were built and trained using the constructed convolutional neural network model,and the recognition accuracy and recognition speed of the convolutional neural network model for different types of defects were analyzed.The impact of freezing different layers on the migration learning results and the law of migration learning on the training of the neural network and on the recognition of defect types were investigated.Finally,based on pixel calibration combining convolutional neural network model with image segmentation and morphological processing,quantitative measurement experiments of defect size and area of HSCs were carried out,and the range and measurement error of quantitative measurement of defect size and area of HSCs were investigated,and the relationship between measurement error and locking frequency was analyzed to realize quantitative measurement of defect size and area of HSCs.