Classification of surface defects in the base metal of pipelines based on complex diagnostics results

Abstract

We discuss issues of classification of operational volumetric and planar surface defects based on the results of complex diagnostics by non-destructive ultrasonic sounding using Rayleigh surface waves generated by an electromagnetic-acoustic transducer and the eddy current method. The paper presents results of feature selection using a variance analysis (ANOVA) and an Extra Trees Classifier algorithm, making it possible to select an optimal eddy current transducer for surface defect classification. The classification of surface defects by the amplitude of ultrasonic and eddy current signals, as well as the phase of the eddy current signal separately is shown to be unambiguous. Models for classifying surface defects as being volumetric or planar are constructed based on statistical methods such as Bayesian inference and the Dempster-Schafer theory. The workability of the constructed classification models is evaluated using metrics such as the Jaccard coefficient and the F1-measure.

Keywords

• surface defects
• ultrasonic testing
• eddy current testing
• complex diagnostics
• joint data evaluation
• machine learning
• bayesian inference
• dempster-schafer theory