Gear Pitting Measurement by Multi-Scale Splicing Attention U-Net

Abstract The judgment of gear failure is based on the pitting area ratio gear. Traditional calculation method mainly rely manual visual inspection. This greatly affected by human factors, and working experience, training degree fatigue detection personnel, so results may be biased. non-contact computer vision measurement can carry out non-destructive testing monitoring under condition machine, has high accuracy. To improve accuracy pitting, a novel multi-scale splicing attention U-Net (MSSA U-Net) explored in this study. An image module first proposed for concatenating output feature maps multiple convolutional layers into map with more semantic information. Then, an applied to select key features map. Given that MSSA adequately uses features, it better segmentation performance irregular small objects than U-Net. On basis designed platform U-Net, methodology measuring proposed. With three datasets, experimental show superior existing typical methods accurately segment different levels due its strong ability. Therefore, effectively determining level pitting.