Application Research Of Infrared Nondestrutive Testing Based On Deep Adversatial Learning

Carbon fiber reinforced polymer(CFRP)has the characteristics of high strength and light weight.Therefore,it is widely used in important fields such as aerospace and military industry.In the production process,the material will be debond and delaminated.For industry fields with high safety requirements,the hidden safety hazards caused by internal defects are undoubtedly fatal.Due to non-contact,fast speed and other advantages,infrared non-destructive testing technology based on optical excitation is a non-destructive testing technology suitable for detecting composite materials.However,affected by background and noise,it is necessary to use the defect detection algorithm to process the original infrared image data.The traditional algorithm has been used to separate the background defect information.But these methods are still limited by the image resolution and detection accuracy.In image processing,deep learning algorithms has been widely used to capture feature and identify target information.In non-destructive testing,due to the noise masking defect information in the data and other reasons,deep learning algorithms have disadvantages such as poor generalization effect,insufficient model capacity,and inability to extract defect features.In this paper,based on the study of deep learning algorithms and optical-pulsed infrared non-destructive testing data,convolutional neural networks and generative adversarial networks have been used as the baseline model.The above problems are improved from the model and architecture.The novel defect detection network suitable has been used for optical pulsed infrared non-destructive testing data.The network improves the detection rate of defects,increases the capacity of the network,and can effectively detect both regular and irregular defects.The main research work of this paper is as follows:1)Two different sets of pulsed optical-pulsed infrared thermal imaging systems were used to perform experiments on a variety of specimens to obtain infrared thermal imaging data.Data analysis has been performed on infrared thermal imaging data.The research is conducted from two main directions: network structure and objective function.The comparision result have been finished under the current popular deep learning semantic segmentation algorithms.The reason and feasibility of combining infrared thermal imaging data with the basic principles of deep learning has been studied.The result on various algorithms has been discussed.2)The infrared thermal imaging dataset has been established and tested on the popular semantic segmentation algorithms.The joint-loss structured deep adversarial algorithms has been proposed for infrared non-destructive detection data.This method is based on the adversarial network framework,and improves the semantic segmentation network as the generation network.The objective function of the new algorithm is improved through the study of the semantic segmentation model and the objective function of the generation of the adversarial model.Finally,the algorithms can detect defect on infrared thermal images of different types of specimens.In order to verify the effectiveness and robustness of the algorithm,F-score has been used as an evaluation index for performance evaluation.Four deep learning semantic segmentation algorithms have been selected as comparison algorithms to compare the algorithm results.The results show that the new algorithm proposed in this paper has better detection performance and robustness.