Automatic Defect Detection And Depth Estimation Based On A Sequence Encoding Technique And Convolutional Neural Network

Non-destructive testing is applied for the detection of material defect or damage without changing the properties of the investigated objects.In many cases,the inspector not only needs to detect the defect in the material,but also needs to quantitatively characterize the defect,such as lateral size,defect depth,then the inspector can take further remedial measures.Nowadays the defect detection and quantitative characterization mainly rely on experienced inspectors,the labor-intensive and costintensive problem hinder the meeting of needs for the needs of modern non-destructive testing.The aim of this thesis is to use infrared thermography nondestructive testing technology and convolutional neural network to achieve automatic defect detection and depth estimation in one step,and to promote the research process of artificial intelligence in the field of automatic defect analysis.In this thesis,infrared thermography nondestructive testing technology is used to obtain the surface temperature-time signals of two kinds of fiber reinforced composites after pulsed thermal excitation.Using the innovative sequence encoding technique proposed in this study,onedimensional temperature-time signals of defect area and non-defect area were converted into two-dimensional features image,and then the defect detection and depth quantitatively was characterized in one step using the powerful image feature extraction capabilities of convolutional neural networks.The proposed algorithm in this thesis was compared with three feedforward neural network based algorithm,and the experimental results show that the model based on sequence encoding algorithm and convolutional neural network performs better in both defect detection and defect depth estimation compared with other models.Finally,the robustness of the proposed model is discussed.The robustness of the model was tested under different conditions by changing the heating energy,sampling frequency and experimental materials.The results show that the model has great robustness under different experimental conditions.37 Figures,2 Tables,and 65 References.