Research On Quantitative Non-destructive Evaluation Based On Multiphysics Imaging And Feature Fusion

Multiphysics imaging and quantitative evaluation of defects under complex surface situations are the main technical issues in the field of nondestructive testing.Aiming at these two key technical issues,this dissertation proposed a feature-level fusion framework based on feature extraction,feature selection,and fusion(FE-S-F).Combining the proposed framework with supervised learning,the pattern deep region learning algorithm was investigated to solve the problem of defect detection and location under a complex surface environment.In addition,combining this framework with unsupervised learning,the deep residual fusion network was proposed to achieve multiphysics imaging and quantitative evaluation under both static and moving conditions.The main research contents and contributions of this dissertation are summarized as follows:1)Aiming at multiphysics imaging and defect location under complex surface situations,a feature-level fusion framework called FE-S-F was proposed.The heuristic fusion strategies designed based on deep learning were discussed for specific case studies.For the case with sufficient samples,data were collected from experiments and augmented to construct the training dataset for supervised `learning.The pattern deep region pattern learning algorithm was investigated for defect location.In the case of limited samples,the deep residual fusion network which was trained on large-scale public datasets in an unsupervised learning fashion was proposed for multiphysics imaging.Through several case studies,this dissertation discussed how to extract and select features according to the characteristics of physical fields,and proved that the proposed framework can be applied to a variety of nondestructive evaluation tasks.2)Combining with the FE-S-F framework,feature extraction methods based on instrument properties and multiphysics were investigated.To tackle the technical issue of the low contrast of defect information in multilayer structures,optical-excitation thermography experiments were carried out on four basalt-carbon hybrid fibre reinforced laminates.Sparse pattern extraction methods were compared on the thermography data.The ultrasonic testing was involved for comparative purposes,which proved the sparse features have enhanced the contrast of defects.To overcome the problem of low SNR caused by the irregular geometry,the low-rank feature was studied for the SNR improvement of a 3D hybrid aluminum-carbon fiber reinforced polymers(CFRP)structure.Pulsed thermography data was collected by long-wave and mid-wave infrared cameras.The low-rank feature is compared with the results of X-ray computed tomography(CT)to verify the effectiveness of this feature for noise suppression.3)Combining the FE-S-F framework with supervised learning,the pattern deep region learning algorithm was designed.To solve the technical issue of the detection and location of micro-cracks on the surface of metallic materials and welding lines,the pattern deep region learning algorithm was adapted to achieve the eddy current pulsed thermography(ECPT)and quantitative evaluation of the defects under the complex surface situation.Using the probability of detection(Po D)as the evaluation index,the pattern deep region learning algorithm was quantitatively compared with other state-ofthe-art crack detection algorithms to verify its effectiveness.4)Based on the FE-S-F framework,the multi-excitation infrared fusion imaging approach was proposed.To tackle the issue of quantitative evaluation of high surface emissivity materials,a multi-excitation infrared fusion imaging system based on opticalexcitation thermography and vibrothermography was designed.Six fibre metal laminates(FMLs)which contain aluminum-basalt fiber reinforced plastic(BFRP)and glass fiber reinforced plastic(GFRP)were evaluated by the multi-excitation infrared fusion imaging approach.Fusion imaging was carried out and the quantitative evaluation was conducted on the deformation area,which verified the effectiveness of this approach on multiphysics imaging and quantitative evaluation of high surface emissivity materials.5)Combining the FE-S-F with unsupervised learning,the deep residual fusion network was proposed.To solve the problem of unbalanced exposure of continuous-wave terahertz(CW THz)dynamic line-scan system,the deep residual fusion network was used for compensating the defect information.An autonomous dynamic line-scan system was designed to inspect the cultural heritage sample marquetry and a famous painting titled Arrangement in Grey and Black No.1.The experimental results verified the effectiveness of the proposed system on quantitative defect evaluation under the moving situation.Real-time imaging was achieved with the speed of 49.2 mm/s.To deal with the challenge of complex structure in plant fibre reinforced polymers(PFRP),the deep residual fusion network was adapted to achieve infrared-terahertz fusion imaging on corresponding depth and three-dimensional imaging on PFRP for the first time.The results were compared with X-ray CT,which verified the 3D imaging technique reached a spatial resolution of 0.5 mm and the effectiveness in the three-dimensional quantitative characterization of the complex structure.These two cases also proved that the deep residual fusion network can be applied for multiphysics imaging and quantitative evaluation in both moving and static situations.