Welcome to our repository dedicated to advancing the field of cross-domain fault diagnosis through transfer learning. Our focus lies in Single-source Unsupervised Domain Adaptation (SUDA), Multi-source Unsupervised Domain Adaptation (MUDA) and Domain Generalization (DG), offering a rich collection of resources and methods.
Our repository includes a diverse range of state-of-the-art methods, each accompanied by relevant publications and direct access to implementation code in models. Highlights include:
- ACDANN - Integrating expert knowledge with domain adaptation for unsupervised fault diagnosis. Published in TIM 2021 | View Code
- ADACL - Adversarial domain adaptation with classifier alignment for cross-domain intelligent fault diagnosis of multiple source domains. Published in Measurement Science and Technology 2020 | View Code
- BSP - Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. Published in ICML 2019 | View Code
- CDAN - Conditional adversarial domain adaptation. Published in NIPS 2018 | View Code
- CORAL - Deep coral: Correlation alignment for deep domain adaptation. Published in ECCV 2016 | View Code
- DAN - Learning transferable features with deep adaptation networks. Published in ICML 2015 | View Code
- DANN - Unsupervised domain adaptation by backpropagation. Published in ICML 2015 | View Code
- MCD - Maximum classifier discrepancy for unsupervised domain adaptation. Published in CVPR 2018 | View Code
- MDD - Bridging theory and algorithm for domain adaptation. Published in ICML 2019 | View Code
- MFSAN - Aligning domain-specific distribution and classifier for cross-domain classification from multiple sources. Published in AAAI 2019 | View Code
- MSSA - A multi-source information transfer learning method with subdomain adaptation for cross-domain fault diagnosis. Published in Knowledge-Based Systems 2022 | View Code
- IRM - Invariant risk minimization. Published in ArXiv 2019 | View Code
- MixStyle - Domain generalization with mixstyle. Published in ICLR 2021 | View Code
- IBN - Two at once: Enhancing learning and generalization capacities via IBN-Net. Published in ECCV 2018 | View Code
- MLDG - Learning to generalize: Meta-learning for domain generalization. Published in AAAI 2018 | View Code
- GroupDRO - Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. Published in ICLR 2020 | View Code
- VREx - Out-of-distribution generalization via risk extrapolation. Published in ICML 2021 | View Code
Our code runs fine with the following prerequisites:
- Python 3 (>=3.8)
- Pytorch (>=1.10)
- Numpy (>=1.21.2)
- Pandas (>=1.5.3)
- tqdm (>=4.46.1)
- Scipy (>=1.10)
You can access our repository either by direct download or using git clone. Here’s how:
- Click on the 'Code' button and select 'Download ZIP'.
- Extract the ZIP file to your desired location.
- Open your command line interface.
- Navigate to the directory where you wish to clone the repository.
- Run the command:
git clone https://github.com/Feaxure-fresh/TL-Bearing-Fault-Diagnosis.gitOur repository supports several public datasets for fault diagnosis, with accompanying loading code. These include:
- CWRU - Case Western Reserve University dataset.
- MFPT - Machinery Failure Prevention Technology dataset.
- PU - Paderborn University dataset.
- XJTU - Xi’an Jiaotong University dataset.
- IMS - Intelligent Maintenance Systems dataset.
- Create a folder named "datasets" in the root directory of the cloned repository.
- Download the desired datasets and place them into this "datasets" folder, follow the steps below:
For analyzing a specific dataset under different working conditions:
- Divide the dataset into separate folders named "condition_0", "condition_1", etc., each representing a unique operational condition.
- Within each "condition_?" folder, create subfolders (with custom names) for different fault categories containing the respective fault data.
- Ensure each 'condition_?' folder contains subfolders with identical names and numbers (indicating the same classes of faults).
For example, for the CWRU dataset:
- Organize the dataset into folders based on motor speed (four speeds as four folders).
- Within each condition folder, categorize data into 9 subfolders for 9 fault classes, such as '7 mil Inner Race fault', '14 mil Inner Race fault', '7 mil Outer Race fault', etc., as detailed in Table XII of this IEEE article.
Example folder structure for CWRU dataset:
.
└── datasets
└── CWRU
├── condition_0
│ ├── ball_07
│ │ └── 122.mat
│ ├── inner_07
│ │ └── 109.mat
│ ...
├── condition_1
│ ├── ball_07
│ │ └── 123.mat
│ ...
├── condition_2
...
For implementing transfer between different datasets:
- Organize the dataset into multiple folders according to fault categories across at least two datasets.
- Maintain consistency in folder names and numbers across all datasets.
For instance, when organizing CWRU and MFPT datasets for one-to-one transfer:
.
└── datasets
├── CWRU
│ ├── inner
| | ├── ***.mat
| | | ***.mat
| | ...
│ ├── normal
│ └── outer
└── MFPT
├── inner
├── normal
└── outer
🌟 Still confused about the dataset setup? Please refer to the sample datasets provided in this repository.
For incorporating other public datasets or your custom datasets, navigate to data_loader/load_methods.py in the repository. Implement your data loading function following this template:
def your_dataset_name(item_path):
# Your code to extract the signal or data from the file
signal = take_out_data_from_file(item_path)
return signalThis process allows for the seamless integration within our framework.
Train models using data from the same dataset but different operational conditions.
Example: Transfer from CWRU operation condition 0 to condition 1.
python train.py --model_name DAN --source CWRU_0 --target CWRU_1 --train_mode single_source --cuda_device 0Example: Transfer from CWRU operation condition 0 and condition 1 to condition 2.
python train.py --model_name MFSAN --source CWRU_0,CWRU_1 --target CWRU_2 --train_mode multi_source --cuda_device 0 Train models using data from different datasets.
Example: Transfer from CWRU to MFPT dataset.
python train.py --model_name DAN --source CWRU --target MFPT --train_mode single_source --cuda_device 0Example: Transfer from CWRU and PU datasets to MFPT dataset.
python train.py --model_name MFSAN --source CWRU,PU --target MFPT --train_mode multi_source --cuda_device 0Load and utilize weights from previously trained models.
Example: Load weights and test on CWRU operation condition 3.
python train.py --model_name MFSAN --load_path ./ckpt/MFSAN/multi_source/**.pth --source CWRU_0,CWRU_1 --target CWRU_3 --cuda_device 0NOTE: The --source flag is not necessary for some models when loading weights for testing. However, for certain models, the number of sources is required to define the model structure, and the specific sources used are not important in this context.
🛠️ For more experimental settings, please modify the arguments in opt.py.
We welcome feedback, inquiries, and suggestions to improve our work. If you encounter any issues with our code or have recommendations, please don't hesitate to reach out. You can contact Jinyuan Zhang via email at feaxure@outlook.com, or alternatively, feel free to post your queries or suggestions in the Issues section of our GitHub repository.
Your support in citing our project when used in your research is highly appreciated. It helps in the recognition and dissemination of our work. Please use the following citation format:
@misc{TL-Bearing-Fault-Diagnosis,
author = {Jinyuan Zhang},
title = {TL-Bearing-Fault-Diagnosis},
year = {2022},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/Feaxure-fresh/TL-Bearing-Fault-Diagnosis}}
}