Ultrasonic detection of white etching defect based on convolution neural network

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