Research On Feature Extraction And Defect Classification Of Guided Wave Defect Signal Based On Sparse Representation

Welding is an indispensable connecting part of pressure vessel,ship,offshore platform and other large structures,which is widely used in industrial production and construction.The work status of weld structure will have a direct impact on the safety of industrial production and people’s normal life.The working environment of weld structure is relatively harsh,and it needs to bear great environmental pressure such as high temperature,high pressure or cold.Long-term exposure to harsh working environment will inevitably lead to weld defects.Therefore,it is of great practical significance for the society to carry out health inspection on weld structure and extract defect information.According to the characteristics of the weld structure,the research shows that the feature guided wave can detect the welding structure more quickly and efficiently.When the weld guided wave detects the weld seam,the echo signal contains all the information of the weld seam.However,due to the dispersion of the guided wave and the interference of the noise,the echo signal is often complex.So,how to extract the defect characteristic information from echo signal is the focus of this paper.With the aim of feature extraction and defect classification of defect echoes in the weld,a novel feature guided wave echo signal processing method about echo signal reconstruction based on sparse representation theory was studied in this paper.The classification method of weld defect based on PCA and SVM was proposed,which realized the effective classification of typical weld defects.The main contents and conclusions are as follows:(1)According to the basic principle of weld guided wave,the basic characteristics of echo signal welding seam were analyzed,and an echo signal processing method based on sparse representation was studied.The transient and sparsity of the weld guided wave echo signal was revealed,which provided a theoretical basis for the sparse representation strategy of the transient components in the echo signal.The key problems about the sparse representation model of the weld guided wave echo signal were summarized as the construction of the over-complete atomic dictionary adapted to the original signal features and the optimization solution of the sparse representation vector.(2)An over-complete Morlet wavelet-based atomic dictionary adapted to the characteristics of the feature guided wave excitation signal and the echo signal was constructed.Firstly,according to the waveform oscillation of the excitation signal modulated by HANNING window was very similar with Morlet wavelet,and both the excitation signal and the echo signal had the same bilateral attenuation characteristics with Morlet wavelet,the Morlet wavelet was selected as the base atom.A correlation filtering method was proposed to obtain the Morlet wavelet atoms which was most similar with the original signal,and an over-complete Morlet wavelet atom dictionary was constructed by the translation expansion of the time parameter.The matching degree between the over-complete atomic dictionary and the signal was guaranteed by this method.(3)The SALAS algorithm was applied to solve the objective function of sparse representation vector.The sparse representation of the transient components in the echo signal and the feature recognition and location of the defect signal were realized.Based on the correlation coefficient and convergence speed,the influence of parameter Lagrange multiplier and penalty factor on the result of sparse representation in split augmented Lagrange contraction algorithm was analyzed,and the selection criterion of parameters was determined.The existence and uniqueness of sparse representation vector was demonstrated.The reliability of this method was verified by plotting the box diagram of sparse echo signal to represent the eigenvalues of the signal in the case of random noise.The application research of simulated signals and the actual acquisition of weld characteristic guided wave detection signals has realized the accurate identification and location of actual weld defects,and verified the validity and applicability of this method for the identification of weld defect signals.(4)The method of weld defect classification based on PCA with SVM was studied,and an effective classification of typical weld defects was realized.Based on the Principle of PCA and SVM,feature extraction of sparse reconstructed defect signal was carried out,and defect feature matrix was composed of sparse components in the sparse representation vector of defect signals.The multi-dimensional feature data was fused and optimized by PCA,and low-dimensional feature matrix which can best represent defect signal characteristics was obtained,which reduced the computational complexity.The effects of different kernel functions and corresponding parameters on SVM classification were analyzed.The dimension-reduced data was used in SVM to train the weld defect classification model,and the effective classification of three kinds of weld states,i.e.hole,crack and no defect,was realized.In summary,a sparse representation model of weld feature guided wave was established in this paper.Then the echo signal was sparse reconstructed,which was very beneficial to the extraction of defect information in weld feature guided wave echo signal,and the defect recognition and location were realized.Based on the reconstructed signals of defects,the precise classification of weld defects was completed by combining PCA and SVM.