KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #ai-agents #developer-tools #codex #claude-code #mcp #rust #deepseek-harness #typescript #ai-agent #cli

SDALR

★ 35 Open GitHub ↗

Source-free domain adaptation based on label reliability for cross-domain bearing fault diagnosis

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Official implementation for **SDALR**

## [**Both Reliable and Unreliable Predictions Matter: Domain Adaptation for Bearing Fault Diagnosis without Source Data**](http://www.baidu.com)

### Framework:

### Prerequisites:
- python == 3.11.10
- pytorch ==2.4.1
- numpy, scipy, sklearn, PIL, argparse, tqdm, openpyxl, collections

### Dataset:

-  Please manually download the datasets [PU](https://pan.baidu.com/s/1d505GjqsmHWlwFG5hb5c3Q?pwd=5m1l), [JNU](https://pan.baidu.com/s/1d505GjqsmHWlwFG5hb5c3Q?pwd=5m1l).  
-  Put the downloaded datasets into `./DATA/` with the following structure:

```
DATA
├── PU_1d_8c_2048
│ ├── data
│ │ ├── K001
│ │ │ ├── sample.txt
│ │ │ └── ...
│ │ ├── K002
│ │ │ ├── sample.txt
│ │ │ └── ...
│ │ └── ...
│ └── label.txt
│
└── JNU_1d_2048_2000
├── data
│ ├── NO
│ │ ├── sample.txt
│ │ └── ...
│ ├── OR
│ │ ├── sample.txt
│ │ └── ...
│ └── ...
└── label.txt
```

- `data/` contains the raw time-series samples organized by fault categories.  
- `label.txt` provides the mapping between samples and their corresponding class labels.

### Experimental Results

#### Diagnosis Accuracy
We evaluate SDALR on the PU and JNU datasets under source-free settings.  
The results show that SDALR consistently outperforms UDA and SFDA baselines.

👉 Detailed per-task results can be found in the [paper](link).

#### Reliability Threshold ($\partial$):
  With $\beta=0.6$, we varied $\partial$ from 0.45 to 0.95.  
  - Too high ($\partial > 0.9$): pseudo-label accuracy fluctuates, unreliable samples increase.  
  - Moderate ($0.4 \leq \partial \leq 0.7$): stable pseudo-label accuracy and diagnostic performance.  
  - Conclusion: $\partial$ governs the trade-off, while voting ensures robustness.

#### Sample Size Analysis:
  With $\beta=0.6$, $\partial=0.6$, we varied the number of target samples.  
  - (a) Accuracy of all pseudo-labels (%)  
  - (b) Fault diagnosis accuracy (%)  
  - Results confirm the scalability and robustness of the method.

### Training

1. ##### Source-free Domain Adaptation (SFDA) on PU dataset
   - Pre-train the source model:
     ```bash
     cd object/
     python src_pretrain_RES_PU.py
     ```
   - Adaptation to the target domain:  
     We provide a shell script for automatic adaptation.  
     The script supports **custom arguments** (e.g., number of epochs, batch size, learning rate) that can be modified directly in the `.sh` file for flexible control.  
     ```bash
     ./choose_s_t_PU.sh
     ```

2. ##### Source-free Domain Adaptation (SFDA) on JNU dataset
   - Pre-train the source model:
     ```bash
     cd object/
     python src_pretrain_RES_JNU.py
     ```
   - Adaptation to the target domain:  
     Similar to PU, the script allows users to **customize various parameters** inside the `.sh` file for automated adaptation.  
     ```bash
     ./choose_s_t_JNU.sh
     ```

3. ##### Ablation Studies
   To reproduce the ablation experiments reported in the paper, simply run:
   ```bash
   ./ablation_PU.sh
   ./ablation_JNU.sh

```

### Contact

- [bd_lab@163.com](mailto:bd_lab@163.com)

SDALR

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README