Feature-selective Clustering And Tensor Analysis For Ultrasonic-based Automatic Defect Identification In CFRP

In recent years,due to the high mechanical stiffness,low density and good chemical stability,carbon fiber reinforced polymer(CFRP)has been rapidly developed and widely used in aerospace,automobile manufacturing,nuclear energy industry,etc.However,affected by manufacturing technology and external impact,the CFRP structure is prone to defects and damage during the manufacture and use.In order to ensure the integrity and reliability of CFRP,it is very important to perform non-destructive testing before and during use.A variety of non-destructive testing(NDT)technologies have been applied to defect identification in CFRP materials,among which ultrasonic testing(UT)is the most popular one due to its high sensitivity,accurate location,wide range of detection,high safety and harmlessness,etc.Affected by signal attenuation,medium absorption,reflection,etc.,raw signals of UT often contain noise,leading to difficulty of defect identification.Various signal and image processing technologies have been utilized for ultrasonic testing,while they highly depend on operators’ experience and expertise and are unable to provide automatic defect detection.Note that the data structure of UT is essentially a three-order tensor,while the traditional methods are usually based on one-dimensional signal or two-dimensional image,which damage the internal structure of high-order data and lead to failure of exploiting hidden information.In order to achieve automatic and intelligent UT,in this thesis,an unsupervised feature-selective clustering method is proposed firstly for data processing of UT.The size,position and depth of the defects can be fully evaluated accurately by adaptively selecting effective feature subset.Considering the limitations of the lower dimensional data analysis methods,a three-order tensor analysis(Tucker3 decomposition)is also presented for ultrasonic detection in this thesis.After Tucker3 decomposition,the defect information can be extracted by a small amount of factor combinations obtained from core tensor,which is further summarized by the leverages.The candidate defective regions are then determined by the leverages,based on which the locations and the shapes of the defects can be identified by clustering analysis.In order to verify the feasibility of the above methods,CFRP specimens with artificial defects were prepared by resin transfer molding method.The experimental data were obtained by ultrasonic phased array detector.The experimental results illustrate the feasibility and effectiveness of the proposed method.