Wind Turbine Gearbox Gear Surface Defect Detection Based on Multiscale Feature Reconstruction

Abstract

The fast and accurate detection of wind turbine gearbox surface defects is crucial for wind turbine maintenance and power security. However, owing to the uneven distribution of gear surface defects and the interference of complex backgrounds, there are limitations to gear-surface defect detection; therefore, this paper proposes a multiscale feature reconstruction-based detection method for wind turbine gearbox surface defects. First, the Swin Transformer was used as a backbone network based on the PSPNet network to obtain global and local features through multiscale feature reconstruction. Second, a Feature Similarity Module was used to filter important feature sub-blocks, which increased the inter-class differences and reduced the intra-class differences to enhance the discriminative ability of the model for similar features. Finally, the fusion of contextual information using the pyramid pooling module enhanced the extraction of gear surface defect features at different scales. The experimental results indicated that the improved algorithm outperformed the original PSPNet algorithm by 1.21% and 3.88% for the mean intersection over union and mean pixel accuracy, respectively, and significantly outperformed semantic segmentation networks such as U-Net and DeepLabv3+.

ESD-YOLOv5: A Full-Surface Defect Detection Network for Bearing Collars

“…Deep-learning-based detection algorithms have strong feature expression ability, generalization ability, and cross-scene ability and thus have been widely applied in the industrial field for defect detection. Examples of such applications include detecting features of textiles [1], light guide plates [2], wire and arc additive manufacturing [3], wind turbine gearbox gears [4], and road damage [5].…”