IMECH-IR  > 流固耦合系统力学重点实验室
融合深度学习的点阵夹层板动力学损伤识别方法研究
Alternative TitleDamage detection for sandwich panels with truss core based on dynamic response combining deep learning
王亚博
Thesis Advisor宋宏伟
2019-05-31
Degree Grantor中国科学院大学
Place of Conferral北京
Subtype硕士
Degree Discipline固体力学
Keyword深度学习 点阵夹层板 损伤识别 卷积神经网络
Abstract

点阵夹层板具有高比强度、高比刚度、隔热隔振等多功能性的特点,在航空航天、汽车等领域中受到越来越多的关注。当被用于极端的环境中时,点阵夹层板由于制备缺陷或损伤可能发生局部屈服,最终导致结构失效和破坏,因此,点阵夹层板的损伤识别问题也逐渐引起很多专家学者的关注。但是由于点阵夹层板结构复杂,可能发生损伤的类型(虚焊、胞元缺失、面板孔洞等)多样且具有分布随机性,解具有不唯一性且识别结果易受到主观因素的影响,其损伤识别具有很大难度。而深度学习方法最大的优势是可以从海量数据中学习并提取特征,不需要人为干预,识别结果客观性强;将深度学习方法应用到点阵夹层板的损伤识别中,将有望突破现有识别方法的局限。但是,要将深度学习应用到点阵夹层板的损伤识别中,就必须要解决数据集构建和模型选择这两个最关键的核心问题,本文的研究工作也主要围绕使用敏感标识量构造胞元缺失缺陷和虚焊缺陷海量数据集和损伤识别的深度网络模型展开。

1 点阵夹层板损伤识别的深度学习模型选择和优化也是本课题的核心。利用损伤敏感标识量表征出胞元缺失缺陷生成数据样本,并制作数据集,分别在ZFVGG-16OHEM等五种模型中进行训练,综合比较训练结果,挑选最适宜的网络模型,然后,通过调整模型中的超参数进行训练,设置最优参数,使网络模型和点阵夹层板损伤数据集具有很好的相容性。

2)构造多重损伤标识量同时对缺陷位置进行表征,解决了单一损伤标识量难以同时识别多种缺陷位置的问题,提高损伤识别方法的准确率。点阵夹层板因构型复杂,可能出现损伤类型较多,单一的损伤标识量并不能识别点阵夹层板中所有类型的损伤,通过使用三种损伤标识量同时对缺陷位置表征,生成丰富的缺陷位置样本,构造出样本多且特征丰富的高质量数据集,充分发挥深度学习特征提取的优势,获得损伤识别准率高的网络模型,最终训练出损伤识别准确度最高的模型,并使用交叉面积(Overlap AreaOLA)方法,解决损伤漏检的问题。

3)优化后的深度学习模型对点阵夹层板中常见的虚焊缺陷同样具有较高的识别准确率。虚焊缺陷是点阵夹层板中最常见的损伤类型,为了研究深度学习模型对虚焊缺陷的识别效果,本文使用柔性梁模拟虚焊缺陷,生成含三种虚焊缺陷(一个虚焊点、两个虚焊点和四个虚焊点)的数据样本,将虚焊缺陷的样本和胞元缺失的样本融合到一起制作用于训练的数据集,通过优化网络参数,使得网络模型和数据集具有更高的适配性,结果表明深度学习方法不仅能识别点阵夹层板中的胞元缺失缺陷,也能识别虚焊缺陷。

Other Abstract

Sandwich panels with truss core (SPTC) have the characteristics of high specific strength, high specific stiffness, thermal insulation and other versatility, and it has received more and more attention in the fields of aerospace, automotive and so on. When used in extreme environments, SPTC is partially yielded due to its own defects, which ultimately leads to structural failure and damage. Therefore, the structural health problems of SPTC have gradually attracted the attention from many experts and scholars. However, due to the complex structure of SPTC, the types of damage (debonding, cell missing, panel voids, etc.) are diverse and random, the solution is not unique and detection result is susceptible to subjective factors. The advantage of the deep learning method is that it can learn and extract features from the massive dates. There is no need for human intervention in the process, and detection results are objective. The application of deep learning method for damage detection of SPTC is expected to break through the limitations of traditional detection method. However, in order to apply deep learning to the damage identification of SPTC, it is necessary to solve the two most critical problems of dataset construction and model selection. The work in this paper is mainly focused on constructing the dataset of cell missing damage and debonding damage based on damage index, and deep learning model for damage identification should be thought.

1) The selection and optimization of deep learning algorithm for debonding damage identification is also the core of this topic. The datasets for cell missing of SPTC are trained in five models such as ZF, VGG-16 and OHEM, and a comprehensive model is found from the most suitable model. Then, by adjusting the hyperparameters the network model of a high degree is finally achieved.

2) The multiple damage indexes are constructed and the damage characteristics are characterized at the same time, which solves the problem that the single damage index is sensitive to different types of damage, and this method improves the accuracy of the damage detection. Due to the complex configuration, SPTC may have more types of damage. The single damage index does not identify all types of damage in SPTC. By using three damage indexes, the damage characteristics are simultaneously characterized and richly. The damage feature samples are generated to construct a high-quality dataset including many samples and rich features, the method fully exploits the advantages of deep learning damage recognition and trains a network model with high probability of damage identification and the method of overlap area (OLA) is applied to solve the problem of damage detection error.

3) The selected deep learning model also has a high recognition accuracy for debonding damage of SPTC. The debonding damage is the most common type of SPTC. In order to study the effects on the deep learning model for detecting debonding damage, in this paper debonding damage is simulating by using the flexible beam, three kinds of debonding damage are generated, the dataset is generated combing cell missing damage samples, and the network parameters are adjusted make the network adaptable with the dataset, and it turns out that the deep learning method can also identify debonding damage SPTC.

Language中文
Document Type学位论文
Identifierhttp://dspace.imech.ac.cn/handle/311007/79121
Collection流固耦合系统力学重点实验室
Recommended Citation
GB/T 7714
王亚博. 融合深度学习的点阵夹层板动力学损伤识别方法研究[D]. 北京. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
王亚博硕士论文 15611550730 (6423KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王亚博]'s Articles
Baidu academic
Similar articles in Baidu academic
[王亚博]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王亚博]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.