# HierMUD: a Hierarchical Multi-task Unsupervised Domain adaptation framework
This is the repository for the paper:
>* Jingxiao Liu, Susu Xu, Mario Bergés, Hae Young Noh. HierMUD: Hierarchical Multi-task Unsupervised Domain Adaptation between Bridges for Drive-by Damage Diagnosis.
[[slides]](docs/slides.pdf)[[paper]](https://arxiv.org/abs/2107.11435)[[video]](docs/video.mp4)
### Description
We introduce HierMUD, a novel approach for multi-task unsupervised domain adaptation. This approach is developed for bridge health monitoring using drive-by vehicle vibrations, but it can be applied to other problems, such as digit recognition, image classification, etc.
![The architecture of our hierarchical multi-task and domain-adversarial learning algorithm. The red and black arrows between blocks represent source and target domain data stream, respectively. Orange blocks are feature extractors, blue blocks are task predictors, and red blocks are domain classifiers.](imgs/arch.png)
In this repository, we demonstrate our approach through two examples:
- A drive-by bridge health monitoring example, which transfers model learned using vehicle vibration data collected from one bridge to detect, localize and quantify damage on another bridge.
- A digit recognition example, which transfers model learned using MNIST data to MNIST-M data and conducts two tasks: odd-even classification and digits comparison.
Note: the drive-by bridge health monitoring experiment involves data that is not publicly available. We will work towards making the experiment replicable without violating data usage policy.
### Code Usage
```
git clone https://github.com/jingxiaoliu/HierMUD.git
cd HierMUD
```
- Run the drive-by bridge health monitoring example with
```
jupyter notebook demo_dbbhm.ipynb
```
- Run the digit recognition example with
```
jupyter notebook demo_mnist.ipynb
```
### Contact
Feel free to send any questions to:
- [Jingxiao Liu](mailto:[email protected]), Ph.D. Candidate at Stanford University, Department of Civil and Environmental Engineering.
### Citation
If you use this implementation, please cite our paper as follows:
```
@misc{liu2021hiermud,
title={HierMUD: Hierarchical Multi-task Unsupervised Domain Adaptation between Bridges for Drive-by Damage Diagnosis},
author={Jingxiao Liu and Susu Xu and Mario Bergés and Hae Young Noh},
year={2021},
eprint={2107.11435},
archivePrefix={arXiv},
primaryClass={cs.AI}
}
```
没有合适的资源?快使用搜索试试~ 我知道了~
桥梁之间的分层多任务无监督域适应用于路过损伤诊断
共68029个文件
png:68003个
gz:4个
py:4个
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
0 下载量 69 浏览量
2022-06-24
14:18:27
上传
评论
收藏 216.7MB ZIP 举报
温馨提示
垂直布局通过驾驶车辆的振动响应来监测桥梁,通过允许每辆车检查多座桥梁,并消除在每座桥上安装和维护传感器的需要,实现了高效和低成本的桥梁维护。然而,许多现有的飞车监控方法都是基于监督式学习模型的,这些模型需要来自每个桥梁的大量标记数据。如果不是不可能的话,获得这些标记的数据是昂贵和费时的。此外,直接将在一座桥上训练的监督式学习模型应用到新的桥上,由于不同桥之间的数据分布会发生变化,所以精度较低。此外,当我们有多个任务(例如,损伤检测、定位和量化)时,分布变化比只有一个任务更具挑战性,因为不同的任务有不同的分布变化和不同的任务难度。 为此,我们引入了 HierMUD,这是第一个层次化多任务无监督域自适应框架,它可以将从一个网桥学到的损伤诊断模型转移到一个新的网桥上,而不需要在任何任务中从新的网桥上贴上任何标签。具体来说,我们的框架以一种对抗的方式学习分层神经网络模型,以提取对于多个诊断任务具有信息性且跨多个桥梁不变的特征。为了匹配多任务的分布,我们设计了一个新的损失函数,该损失函数基于一个新的可证明的泛化风险约束,自适应地为具有更多移动分布的任务分配更高的权重。为了学习具有不同任务难度的多
资源推荐
资源详情
资源评论
收起资源包目录
桥梁之间的分层多任务无监督域适应用于路过损伤诊断 (68029个子文件)
train-images-idx3-ubyte.gz 9.45MB
t10k-images-idx3-ubyte.gz 1.57MB
train-labels-idx1-ubyte.gz 28KB
t10k-labels-idx1-ubyte.gz 4KB
Icon
0B
Icon
0B
desktop.ini 136B
desktop.ini 136B
demo_mnist.ipynb 273KB
demo_dbbhm.ipynb 8KB
README.md 2KB
video.mp4 39MB
slides.pdf 4.05MB
arch.png 158KB
00012974.png 3KB
00025840.png 3KB
00039745.png 3KB
00012750.png 3KB
00040561.png 3KB
00017573.png 3KB
00015087.png 3KB
00002934.png 3KB
00053364.png 3KB
00051088.png 3KB
00057490.png 3KB
00055142.png 3KB
00030492.png 3KB
00044004.png 3KB
00012720.png 3KB
00044144.png 3KB
00003776.png 3KB
00023965.png 3KB
00000945.png 3KB
00037244.png 3KB
00010424.png 3KB
00008126.png 3KB
00052208.png 3KB
00003161.png 3KB
00030461.png 3KB
00030154.png 3KB
00006317.png 3KB
00039751.png 3KB
00050657.png 3KB
00028338.png 3KB
00036378.png 3KB
00055289.png 3KB
00029867.png 3KB
00027108.png 3KB
00005786.png 3KB
00014325.png 3KB
00040983.png 3KB
00050658.png 3KB
00016867.png 3KB
00054097.png 3KB
00029879.png 3KB
00011638.png 3KB
00050910.png 3KB
00048114.png 3KB
00008195.png 3KB
00007125.png 3KB
00032251.png 3KB
00049522.png 3KB
00053296.png 3KB
00017142.png 3KB
00004855.png 3KB
00015437.png 3KB
00052872.png 3KB
00000054.png 3KB
00027710.png 3KB
00001966.png 3KB
00017854.png 3KB
00053242.png 3KB
00007295.png 3KB
00058089.png 3KB
00057725.png 3KB
00031621.png 3KB
00002691.png 3KB
00039156.png 3KB
00002701.png 3KB
00004984.png 3KB
00012313.png 3KB
00021659.png 3KB
00032811.png 3KB
00022379.png 3KB
00023381.png 3KB
00050131.png 3KB
00013230.png 3KB
00008936.png 3KB
00027122.png 3KB
00011781.png 3KB
00024016.png 3KB
00035098.png 3KB
00004365.png 3KB
00055038.png 3KB
00023531.png 3KB
00003596.png 3KB
00048197.png 3KB
00002833.png 3KB
00004322.png 3KB
00055807.png 3KB
共 68029 条
- 1
- 2
- 3
- 4
- 5
- 6
- 681
资源评论
科研吧
- 粉丝: 11
- 资源: 218
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功