Ultrasonic detection of white etching defect based on convolution neural network

Abstract

<sec>Unlike classical defects formed by rolling contact fatigue, white etching defect (WED) including white etching area and white etching crack will cause surface to spall in the early stage and the service life to shorten seriously. Located in the subsurface of bearings, the tiny size WED is difficult to detect by conventional ultrasonic methods. The root cause of WED generation remains unclear. It is time consuming and expensive to prepare samples during the evolution of such defects. For characterizing the WED at early stage, five evolving states concerning the existing microscopic information are established in this paper. The immersion ultrasonic inspection process is simulated based on <i>k</i>-space pseudo spectrum method.</sec><sec>For the later evolutionary stage with crack, the bearing can be simplified into a homogeneous three-layer model by ignoring the internal grain structure. The crack depth is obtained by using the ultrasonic reflection coefficient amplitude spectrum (URCAS), with an error of 1.5%. For other states without crack, the spectrum characteristic is no longer evident with slight acoustic impedance difference between layers. The polycrystalline structure on a microscale is thus realized based on Voronoi diagram, from which the grain induced backscattering can be used to amplify the microstructure variations at different stages. The backscattering signal is influenced by the grain size and detection frequency from the simulation. Since a direct comparison of backscattering information among evolutionary stages is difficult, the five different evolutionary stages of WED are recognized with the help of deep learning. The received waveform is transformed into a time-frequency map by short-time Fourier transform. Based on RESNET network structure, the results show that the train accuracy and validation accuracy reach 92% and 97% respectively. This study provides a sound way to characterize WED, which is conducive to early failure prediction and residual life evaluation.</sec>