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

As the key connection structure of large-scale medium and thick plates,long welds are widely used in industrial manufacturing fields such as deep-sea ships.The structural health of long welds has a direct impact on life and property in production.However,due to welding errors and weld structures under extreme environments for a long time,the weld defects are prone to form.Therefore,it is of great significance for safe production and life to carry out regular health detection on welds,extract their defect characteristics,and other information to deal with them in time.The ultrasonic guided wave detection method is suitable for the detection of large-scale weld structure defects in platforms such as offshore platforms with its advantages of long-distance propagation,small attenuation,high accuracy,and high efficiency.However,due to the impact of inherent characteristics such as dispersion,multi-mode and the interference of background noise,the wave packets of the echo signal will be overlapped,and the defect information will be submerged in the complex wave packets.In this paper,to extract the defect components in the complex wave packet and achieve defect classification,a method of sparse-based defect detection of weld feature guided waves with a fusion of shear wave characteristics is proposed,and defects are classified combined with Fisher linear discriminant analysis.The specific research content and effective conclusions are as follows:（1）Combining theoretical research and simulation analysis to explore the formation mechanism and inherent properties of the weld feature guided wave signal,this paper analysis the characteristics of the defect echo signal and proposes a basic strategy for the sparse feature extraction of guided wave signals and the construction method of the over-complete atom dictionary with a fusion of shear wave characteristics.Results show that the echo signal of weld defects has the characteristics of"transient"and"sparseness"in the time domain,so the signal can be processed by sparse representation.Two key issues in signal processing under the sparse framework are clarified,that is,the construction of an"appropriately sparse"over-complete atom dictionary and the solution of sparse representation coefficients.Combining the excitation signal waveform,propagation distance,dispersion characteristics,and dispersion,and other prior information to estimate the echo signal.This paper selects the estimated signal on each propagation distance is regarded as an atom of the over-complete atom dictionary.The distance step is set appropriately small to ensure the accuracy of the feature recognition of the reconstructed signal.（2）Adopting the l₁ norm constrained basis pursuit denoising model to convert the non-determined equation constrained by the l₀ norm into a linear solution problem,and combined with the SALSA algorithm to solve the sparse coefficients,this paper achieves the results of the sparse reconstruction of the defect echo signal.The finite element simulation signal is analyzed to verify the feasibility of the proposed method,and the anti-noise ability of the proposed method is analyzed under different noise levels to verify the effectiveness of the proposed method.Based on the location error and the difference in signal-to-noise ratio between the reconstructed signal and the original signal,the optimal value of the penalty factor in the SALSA algorithm is determined;the existence and uniqueness of the sparse representation coefficient solution are analyzed;the proposed method is successful applied to the processing of guided wave signals of weld defects,and it realizes defect location and defect feature information acquisition,which lays a foundation for the identification of different types of defects.（3）The classification method of weld defects based on Fisher linear discriminant analysis model is proposed,which realizes the recognition of different types of weld defects.Based on the sparse feature extraction of the guided wave signal,according to the five dimensionless parameters of signal entropy,dispersion coefficient,margin factor,skewness coefficient,and kurtosis coefficient,combined with the sparse component of the defect signal sparse solution coefficient,the feature set of the defect signal is established.Then,the feature set is input into the Fisher linear discriminant analysis model,which realizes the recognition of different types of defects and the estimation of model performance by cross-validation.In this paper,the sparse feature extraction of guided wave signals is researched under the sparse representation framework,and a construction method of an over-complete atom dictionary with a fusion of shear wave characteristics is proposed.Based on the sparse feature extraction,different types of welds are identified.The classification and identification of weld defects are of great significance to the health detection and early maintenance of weld structures.