Deformation Feature Extraction and Double Attention Feature Pyramid Network for Bearing Surface Defects Detection

In vision-based bearing defect detection, random shapes and multiscale defects can result in inaccurate feature extraction, and complex backgrounds can interfere with defect localization. Therefore, a novel bearing defect detection network (BDDN) is proposed in this article, aiming to improve the accuracy of bearing defect detection, which consists of deformable ResNet50 network (DRN) and double attention feature pyramid network (DA-FPN). In the DRN, it is found that using deformable convolution for stages 1–3 of ResNet50 is able to extract features that are closer to the actual defect regions through rich experiments. The DA-FPN is employed to reduce the interference of complex backgrounds for defect detection, which consists of a feature-enhanced attention module (FEAM) and a feature-communication attention module (FCAM). The FEAM obtains more refined high-level semantic information in channel feature extraction, while the FCAM efficiently fuses defect features across different levels. Experimental results demonstrate that the BDDN effectively detects bearing cover surface defects, achieving a mean average precision of 87.9%. In addition, it outperforms several defect detection models on the public Northeastern University dataset (NEU-DET) in terms of accuracy.