🌐 Domain-Adversarial Neural Network (DANN) for Domain Adaptation - 1D Signal

This project implements a DANN (Domain-Adversarial Neural Network) to perform robust fault classification under domain shift using the bearing dataset.

---

📁 Project Structure

dann_pipeline/
├── main.py              # Entry point (load, train, evaluate)
├── config.py            # Argument parser
├── dataset.py           # Source/target dataloaders
├── model.py             # DANN architecture (feature extractor + classifiers)
├── train.py             # Adversarial training routine
├── test.py              # Final evaluation on target domain
├── utils.py             # Optional tools (e.g. plotting)
├── data/                # .npy input files
└── result/              # model.pt, logs, loss curves

---

📥 Dataset

This project uses two publicly available bearing fault datasets adapted for domain adaptation experiments:

• CWRU Bearing Dataset – Provided by Case Western Reserve University
  🔗 https://engineering.case.edu/bearingdatacenter/download-data-file


• IMS Bearing Dataset – Provided via Data.gov
  🔗 https://catalog.data.gov/dataset/ims-bearings

For signal preprocessing and conversion into .npy format, refer to the preprocessing code in this repository:
🔧 https://github.com/97yong/signal-fault-classification

Preprocessed data is saved in the following format as .npy:

data/
├── X_source.npy
├── Y_source.npy
├── X_target.npy
├── Y_target.npy
├── X_target_test.npy
└── Y_target_test.npy

---

⚙️ Training Options

You can configure training parameters via config.py or pass them through opt.

Argument	Description	Default
--epochs	Number of training epochs	10
--lr	Learning rate	1e-4
--batch_size	Mini-batch size	64

---

🧠 Model Overview

DANN consists of:

• Feature Extractor – 1D CNN layers
• Label Classifier – Predicts class labels for source domain
• Domain Classifier – Predicts domain (source/target) using GRL (gradient reversal)

Training is done with adversarial loss to align the feature distributions.

---

🚀 How to Run

pip install numpy torch scikit-learn tqdm matplotlib

python main.py

This will:

1. Load source/target domain data
2. Train a domain-adversarial model
3. Evaluate performance on target test data

---

Views

📚 Reference

Ganin, Y., & Lempitsky, V. (2015).
"Unsupervised Domain Adaptation by Backpropagation."
In Proceedings of the 32nd International Conference on Machine Learning (ICML), 1180–1189.
🔗 arXiv:1409.7495

---