Characteristic analysis and pattern recognition of Arc sound under typical penetration status in MIG welding

Abstract

Aiming at proposing an online monitoring method of penetration status in MIG welding, audible arc sound signal under partial penetration, unstable penetration, full penetration and excessive penetration in the course of flat butt welding with spray transfer was collected, processed and analyzed. And then 11 characteristic parameters, which can characterize weld penetration status from the perspectives on time, frequency, cepstrum and geometry-domains, were extracted by using wavelet de-noising and short-time windowing. At last, 8-dimensional eigenvector with most information of penetration status were re-synthesized with the help of feature-level parameter fusion technology of principal component analysis (PCA). Thereby, taking 8-dimensional eigenvector as input and viewing four penetration status as export. network models for identifying penetration status about BP and RBF were established. The application of test models proved that both constructed networks could realize the online recognition of penetration status. Moreover, the accuracy rate in RBF network was 6.25% more than BP, and arrived at 91.25%.