航空学报 > 2025, Vol. 46 Issue (21): 532355-532355   doi: 10.7527/S1000-6893.2025.32355

中国飞机强度研究所建所 60 周年专刊

基于Transformer的航空结构表面裂纹智能追踪方法

李嘉欣1,2(), 吕帅帅1,2, 王叶子1,2, 杨宇1,2, 李梓悦1   

  1. 1.中国飞机强度研究所,西安 710065
    2.强度与结构完整性全国重点实验室,西安 710065
  • 收稿日期:2025-06-03 修回日期:2025-07-03 接受日期:2025-08-11 出版日期:2025-08-29 发布日期:2025-08-28
  • 通讯作者: 李嘉欣 E-mail:1056900948@qq.com
  • 基金资助:
    国家级项目

Transformer-based intelligent tracking method of aviation structure surface cracks

Jiaxin LI1,2(), Shuaishuai LYU1,2, Yezi WANG1,2, Yu YANG1,2, Ziyue LI1   

  1. 1.Aircraft Strength Research Institute of China,Xi’an 710065,China
    2.National Key Laboratory of Strength and Structural Integrity,Xi’an 710065,China
  • Received:2025-06-03 Revised:2025-07-03 Accepted:2025-08-11 Online:2025-08-29 Published:2025-08-28
  • Contact: Jiaxin LI E-mail:1056900948@qq.com
  • Supported by:
    National Level Project

摘要:

基于深度卷积网络的语义分割模型在结构损伤检测领域展现出了良好的应用效果,但在面向航空结构损伤检测时,由于裂纹通常在图像中的占比很小,多层的卷积、池化操作会导致裂纹信息丢失,严重降低分割精度。因此,对基于Transformer的语义分割模型开展研究,设计了适用于航空结构表面损伤检测的裂纹智能追踪通用模型TICT来实现对裂纹的精准分割和智能追踪。首先,使用自适应动态图像块划分机制将图像分割成不同大小、重叠程度的图像块;其次,将图像块输入基于Transformer的编码器中,提取包含裂纹图像上下文信息、局部细节信息的多尺度特征;然后,使用一个轻量级的多层感知机、注意力模块作为解码器生成裂纹掩码图像;最后,通过图像形态学操作对掩码图像中的裂纹连通域进行修正,并映射回原始图像得到精确的裂纹分割区域;通过对疲劳试验过程中实时采集图像重复上述操作,即可实现对裂纹的自动化持续追踪。在金属元件、全机疲劳试验的裂纹图像数据集上对TICT模型进行训练、测试,TICT模型在多种金属元件、全机结构裂纹图像测试集上平均交并比(mIoU)达到了78.31%,证明了TICT模型对各种结构构型、背景复杂、特征微小的航空结构表面裂纹均能够实现精准分割,具有良好的泛化性、鲁棒性。

关键词: 裂纹追踪, Transformer, 计算机视觉, 语义分割, 结构健康监测

Abstract:

Semantic segmentation models based on deep convolutional networks have shown good performance in structural damage detection. However, when it comes to aircraft structural damage detection, cracks usually occupy a small proportion of the image, and the multi-layer convolution and pooling operations can lead to the loss of crack information, thereby seriously reducing the segmentation accuracy. Consequently, this research is conducted on Transformer-based semantic segmentation models, and de-signs a Transformer-based Model for Intelligent Crack Tracking (TICT) for aeronautical structural surface damage detection, aiming to achieve precise segmentation and intelligent tracking of cracks. To start with, an adaptive dynamic patch partitioning mechanism is employed to divide the image into patches of different sizes with varying degrees of overlap. Next, these patches are fed into a Transformer-based encoder to extract multi-scale features containing both the contextual and local details of the crack image. Then, a lightweight multi-layer perceptron along with attention modules is utilized as a decoder to generate a crack mask image. After that, morphological operations are performed on the mask image to correct the connected regions of cracks and map them back to the original image, thus obtaining the exact crack areas. By repeating the aforementioned procedure on the images collected in real time during fatigue tests, automated and continuous tracking of cracks can be realized. The TICT model is trained and tested on datasets of fatigue test images of metal components and entire aircraft. It achieves an Mean Intersection Over Union (mIoU) of 78.31% on the test sets of crack image of various metal components and full-scale aircraft structures, which demonstrates that the TICT model can accurately segment surface cracks in aviation structures with various structural configurations, complex backgrounds, and tiny features, exhibiting good generalization and robustness.

Key words: crack tracking, Transformer, computer vision, semantic segmentation, structural health monitoring

中图分类号: