摘要： 水电枢纽在长期运行过程中容易受水流侵蚀、应力变化等因素影响，导致形成裂缝、渗漏、脱落、露筋等缺陷，造成重大安全隐患。目前，水电枢纽缺陷识别主要依靠人工巡检，存在效率低、风险高等问题。提出一种水电枢纽缺陷识别方法，基于完全自注意力机制构建缺陷识别网络，以提高网络捕捉长距离全局信息的能力和缺陷识别精度。设计2个同尺寸的自注意力编码器分支，通过双分支结构完成不同尺度自注意力计算，从而提取多尺度缺陷特征，增强全局语义表达能力。构建一个基于类别向量的自注意力混合融合模块，并对2条分支的多尺度特征进行融合，以有效应对水电枢纽缺陷图像尺度变化大、形态多样等问题。在四川某水电站枢纽缺陷数据集上的实验结果表明，该方法宏查准率可达98.87%，缺陷识别效果优于SVM、ResNet-50、MobileNet v3等方法。

关键词: 水电枢纽, 缺陷识别, 自注意力机制, 多尺度特征, 自注意力融合

Abstract: In the long-term operation, hydropower projects are vulnerable to water erosion, stress changes and other factors, resulting in cracks, leakage, falling off, exposed reinforcement and other defects, which can easily cause major safety hazards.Currently, defect identification in hydropower projects mainly depends on manual inspection;however, this method has the problems of low efficiency and high risk.A defect identification method for hydropower project is proposed.Based on a complete self-attention mechanism, a defect identification network is constructed to improve the ability of the network to capture long-distance global information and the accuracy of defect identification.Two branches of the self-attention coder with the same size are designed, and the self-attention calculation of different scales is completed through a double branch structure, to extract the multi-scale defect features and enhance the global semantic expression ability.A self-attention hybrid fusion module based on category vector is constructed, and the multi-scale features of the two branches are fused to effectively deal with the problems of large-scale changes and diverse forms of the defect image of a hydropower project.The experimental results on the defect dataset of a hydropower station in Sichuan Province show that the macro precision of this method can reach 98.87%, and the defect recognition effect is better than SVM, ResNet-50, MobileNet v3 and other methods.

Key words: hydropower complex, defect recognition, self-attention mechanism, multi-scale features, self-attention fusion

中图分类号: 

• TP319.41