A Novel Method for Segmentation and Detection of Weld Defects in UHV Equipment Based on Multiscale Feature Fusion

Abstract

A novel method for detecting weld defects in ultra-high voltage (UHV) equipment is present by combining unimodal semantic segmentation with X-ray imaging. The approach begins by employing a deep neural network to extract weak weld features from X-ray images. A channel attention module is introduced to balance the importance of different feature weights, enhancing the network’s ability to focus on key features. An atrous spatial pyramid pooling module is then utilized to expand the receptive field, effectively leveraging the spatial hierarchical information within the X-ray images. Additionally, a multi-scale feature fusion module is applied to automatically learn feature relationships, capturing semantic information at various scales, which significantly improves the distinction between defective and normal weld regions. The method’s effectiveness is validated through repeated experiments on the GDXray weld dataset and a self-constructed UHV weld dataset. Quantitative comparisons demonstrate that the proposed method significantly enhances the segmentation accuracy of weld defects in UHV equipment, providing a valuable tool for technicians in the field of weld non-destructive testing.