# 信号目标智能识别系统

**Repository Path**: Hannah_M/Intelligent-Signal

## Basic Information

- **Project Name**: 信号目标智能识别系统
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MulanPSL-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-05-01
- **Last Updated**: 2026-05-01

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# RadioML 2016 High-SNR CNN Classification

This project trains a PyTorch 1D CNN on RadioML 2016 samples where `SNR > 5 dB`.
It reads the common `RML2016.10a_dict.pkl` / `RML2016.10b.dat` pickle format,
then writes training curves, classification metrics, confusion matrices, and
accuracy grouped by SNR.

## Environment

Use the requested Anaconda environment:

```powershell
conda run -n py3.9 python --version
```

The script uses PyTorch, NumPy, pandas, scikit-learn, matplotlib, and seaborn.
By default it applies per-sample IQ power normalization and uses GroupNorm in
the CNN, which is more stable than BatchNorm for quick CPU smoke tests.

## Dataset Location

Place the RadioML 2016 dataset here:

```powershell
D:\Project\SC_demo\data\RML2016.10a_dict.pkl
```

If you use RadioML 2016.10B, this path is also fine:

```powershell
D:\Project\SC_demo\data\RML2016.10b.dat
```

## Run

RadioML 2016.10A:

```powershell
cd D:\Project\SC_demo
conda run -n py3.9 python train_high_snr_cnn.py `
  --data D:\Project\SC_demo\data\RML2016.10a_dict.pkl `
  --out runs\high_snr_cnn_2016 `
  --epochs 20 `
  --batch-size 512
```

RadioML 2016.10B:

```powershell
conda run -n py3.9 python train_high_snr_cnn.py `
  --data D:\Project\SC_demo\data\RML2016.10b.dat `
  --out runs\high_snr_cnn_2016b `
  --epochs 20 `
  --batch-size 512
```

For a quick smoke test:

```powershell
conda run -n py3.9 python train_high_snr_cnn.py `
  --data D:\Project\SC_demo\data\RML2016.10a_dict.pkl `
  --out runs\smoke_high_snr_2016 `
  --epochs 3 `
  --max-samples 12000
```

## Outputs

The output folder contains:

- `best_model.pt`: best checkpoint selected by validation accuracy.
- `training_history.csv`, `accuracy_curve.png`, `loss_curve.png`.
- `classification_report.csv`: per-class precision, recall, and F1.
- `confusion_matrix.csv`, `confusion_matrix.png`,
  `confusion_matrix_normalized.png`.
- `accuracy_by_snr.csv`, `accuracy_by_snr.png`: performance for each retained
  high-SNR value.
- `summary.json`: overall accuracy, macro F1, weighted F1, and best/worst SNR.

## How To Read The Results

Use `overall_accuracy` and `weighted_f1` for headline performance. Use
`macro_f1` to judge whether harder modulation classes are being handled fairly.
The normalized confusion matrix is the best place to inspect which modulations
are confused, especially pairs such as `QAM16/QAM64`, `AM-DSB/WBFM`, or
`QPSK/8PSK`. For the `SNR > 5 dB` subset, the retained SNR values are usually
`6, 8, 10, 12, 14, 16, 18 dB`.

## Frontend Demo

The demo uses a separated frontend and backend:

- Backend: `backend/main.py`, FastAPI inference API.
- Frontend: `frontend/index.html`, static upload page.
- Model: `runs/high_snr_cnn_2016/best_model.pt`.

Start the backend:

```powershell
cd D:\Project\SC_demo
conda run -n py3.9 uvicorn backend.main:app --host 127.0.0.1 --port 8000
```

Start the frontend static server in another terminal:

```powershell
cd D:\Project\SC_demo\frontend
conda run -n py3.9 python -m http.server 5500 --bind 127.0.0.1
```

Then open `http://127.0.0.1:5500` in the browser.

Supported upload file formats are `.npy`, `.csv`, `.txt`, `.dat`, and `.json`.
The uploaded file should contain one IQ sample shaped `(2, 128)`, `(128, 2)`,
or a flat list of `256` values.

Create a sample upload file from the dataset:

```powershell
conda run -n py3.9 python scripts\create_sample_signal.py `
  --data D:\Project\SC_demo\data\RML2016.10a_dict_optimized.pkl `
  --out D:\Project\SC_demo\data\sample_signal.npy `
  --mod BPSK `
  --snr 18
```