Quantitative ultrasonic testing for near-surface defects of large ring forgings using feature extraction and GA-SVM

• Published in: Applied acoustics Vol. 173; p. 107714
• Main Authors: Guan, Shanyue, Wang, Xiaokai, Hua, Lin, Li, Li
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2021
• Subjects: Feature extraction; GA-SVM; Near-surface blind; Ultrasonic testing; Feature extraction; Ultrasonic testing; Near-surface blind; GA-SVM
• ISSN: 0003-682X; 1872-910X
• DOI: 10.1016/j.apacoust.2020.107714

•The size (0-1.5mm) and depth (0-2mm) of near-surface defects were divided into four equal sections to lock in a smaller range for quantitative testing.•The pulse width and amplitude of the interface echo were selected as the time-domain features, the main frequency and resonance frequency features were extracted by fast Fourier transform (FFT).•Intrinsic time-scale decomposition (ITD) was used to obtain statistical features of high-frequency PR components and low-frequency PR components energy.•Intelligent identification size and depth small-range of near-surface defects based on feature extraction and GA-SVM. Water immersion ultrasonic testing technology is widely used in non-destructive testing of large ring forgings, but near-surface defects are difficult to identify in near-surface blind zone, because the defect echo always overlaps with interface echo. In this article, a method combining signal feature extraction with genetic algorithm optimization support vector machine (GA-SVM) was proposed to realize quantitatively testing of near-surface defects. Firstly, the near-surface artificial defects were machined on the test specimens from the large ring forgings, and a total of 160 signal samples were obtained by water immersion ultrasonic experiments. Then the time-domain features of ultrasonic signals were extracted, the spectrum features were obtained by fast Fourier transform. And the ultrasonic signals were processed by intrinsic time-scale decomposition, and the statistical features of the rotating components with different frequencies were extracted. Finally, three kinds of neural network classifiers were used to identify the size and depth of defects by the features database. The experimental results showed that the identification error of the GA-SVM classifier with the different number of sample databases was always minimal. The ultrasonic testing experiment of near-surface notch defects further verified the correctness and feasibility of the GA-SVM classifier.