weswebunited/rtl-ml-dataset

---
license: mit
task_categories:
- time-series-forecasting
- audio-classification
language:
- en
tags:
- rf-signals
- radio
- rtl-sdr
- signal-processing
- machine-learning
- telecommunications
- software-defined-radio
pretty_name: RTL-ML RF Signal Classification Dataset
size_categories:
- 100K<n<1M
---

# RTL-ML Dataset

## Dataset Summary

This dataset contains 240 validated RF signal samples captured using an RTL-SDR Blog V4 dongle. It's designed for training machine learning models to classify common RF signals.

**Total Size:** 1.9 GB
**Samples:** 240 (30 samples × 8 classes)
**Format:** NumPy arrays (.npy files)
**Sample Rate:** 1.024 MSPS
**Sample Duration:** 1 second per capture

## Signal Classes

| Class | Frequency | Count | Description |
|-------|-----------|-------|-------------|
| ADS_B | 1090 MHz | 30 | Aircraft transponder signals |
| APRS | 144.39 MHz | 30 | Amateur radio position reporting |
| FM_broadcast | 88-108 MHz | 30 | Commercial FM radio stations |
| ISM_sensors | 433.92 MHz | 30 | Wireless sensors & remote controls |
| NOAA_APT | 137.5 MHz | 30 | Weather satellite imagery |
| NOAA_weather | 162.4 MHz | 30 | Weather radio broadcasts |
| noise | Various | 30 | Background RF noise baseline |
| pager | 931.9375 MHz | 30 | POCSAG pager transmissions |

## Validation Metrics

- **ISM Sensors:** 20.6x burst ratio (strong on/off keying)
- **NOAA Weather:** 14.4 dB SNR (clear signal)
- **Pager/APRS:** 12.7 dB SNR (good quality)
- **Model Accuracy:** 87.5% on test set

## Usage

```python
from huggingface_hub import snapshot_download
import numpy as np

# Download entire dataset
dataset_path = snapshot_download(
repo_id="TrevTron/rtl-ml-dataset",
repo_type="dataset"
)

# Load a sample
sample = np.load(f"{dataset_path}/datasets_validated/ADS_B_0.npy")
print(f"Signal shape: {sample.shape}") # (1048576,) complex64
```

## Dataset Structure

```
rtl-ml-dataset/
└── datasets_validated/
├── ADS_B_0.npy ... ADS_B_29.npy (30 files)
├── APRS_0.npy ... APRS_29.npy (30 files)
├── FM_broadcast_0.npy ... _29.npy (30 files)
├── ISM_sensors_0.npy ... _29.npy (30 files)
├── NOAA_APT_0.npy ... NOAA_APT_29.npy (30 files)
├── NOAA_weather_0.npy ... _29.npy (30 files)
├── noise_0.npy ... noise_29.npy (30 files)
└── pager_0.npy ... pager_29.npy (30 files)
```

Each `.npy` file contains:
- **Shape:** (1048576,) - 1 second @ 1.024 MSPS
- **Dtype:** `complex64` (I/Q samples)
- **Size:** ~8.4 MB per file

## Hardware

- **SDR:** RTL-SDR Blog V4 ($39.95)
- **Computer:** Indiedroid Nova 16GB ($179.95)
- **Antenna:** Telescopic dipole (included)

## Model Performance

When trained with Random Forest (100 trees):
- **Overall Accuracy:** 87.5%
- **Perfect Classes:** ADS-B, FM, ISM, NOAA APT, Weather, Pager (100%)
- **Challenging:** APRS ↔ Noise confusion (sparse packets)

## Citation

```bibtex
@misc{rtl-ml-dataset,
author = {TrevTron},
title = {RTL-ML Dataset: Validated RF Signal Captures},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/TrevTron/rtl-ml-dataset}}
}
```

## License

MIT License - Free for commercial and non-commercial use.

## Related

- **Code:** [github.com/TrevTron/rtl-ml](https://github.com/TrevTron/rtl-ml)
- **Blog:** [unland.dev](https://unland.dev) (coming soon)
- **Hardware Guide:** [Indiedroid Nova Setup](https://github.com/TrevTron/rtl-ml/blob/main/docs/HARDWARE_SETUP.md)

## Contributions

Captured in Temecula, CA (Southern California) using:
- Clear line of sight to multiple signal sources
- Validated with spectral analysis and manual inspection
- All samples meet minimum SNR requirements (>10 dB for modulated signals)

For questions or improvements, see the [GitHub repository](https://github.com/TrevTron/rtl-ml).

许可证：MIT
任务类别：
- 时间序列预测
- 音频分类
语言：
- 英语
标签：
- 射频（RF）信号
- 无线电
- RTL-SDR
- 信号处理
- 机器学习
- 电信
- 软件定义无线电
官方名称：RTL-ML射频信号分类数据集
样本量区间：10万 < 样本量 < 100万

# RTL-ML 数据集

## 数据集概览
本数据集包含240份经过验证的射频（RF）信号样本，均通过RTL-SDR Blog V4接收棒采集而成，旨在用于训练机器学习模型以实现常见射频信号的分类任务。

**总容量：** 1.9 GB
**样本数量：** 240（8个类别，每类30份样本）
**存储格式：** NumPy数组（.npy文件）
**采样率：** 1.024 MSPS（兆样本每秒）
**单次采集时长：** 1秒

## 信号类别
| 类别 | 工作频率 | 样本数 | 描述 |
|----------------|-------------------|--------|--------------------------|
| ADS_B | 1090 MHz | 30 | 航空器应答机信号 |
| APRS | 144.39 MHz | 30 | 业余无线电位置报告信号 |
| FM_broadcast | 88-108 MHz | 30 | 商用调频广播电台信号 |
| ISM_sensors | 433.92 MHz | 30 | 无线传感器与遥控器信号 |
| NOAA_APT | 137.5 MHz | 30 | 气象卫星图像信号 |
| NOAA_weather | 162.4 MHz | 30 | 气象无线电广播信号 |
| noise | 任意频段 | 30 | 背景射频噪声基准样本 |
| pager | 931.9375 MHz | 30 | POCSAG寻呼机传输信号 |

## 验证指标
- **ISM传感器类：** 20.6倍突发比（强开关键控调制）
- **NOAA气象类：** 14.4 dB信噪比（清晰信号）
- **寻呼机/APRS类：** 12.7 dB信噪比（优质信号）
- **模型测试集准确率：** 87.5%

## 使用方法
python
from huggingface_hub import snapshot_download
import numpy as np

# 下载完整数据集
dataset_path = snapshot_download(
repo_id="TrevTron/rtl-ml-dataset",
repo_type="dataset"
)

# 加载单份样本
sample = np.load(f"{dataset_path}/datasets_validated/ADS_B_0.npy")
print(f"信号形状：{sample.shape}") # (1048576,) complex64

## 数据集结构

rtl-ml-dataset/
└── datasets_validated/
├── ADS_B_0.npy ... ADS_B_29.npy （共30个文件）
├── APRS_0.npy ... APRS_29.npy （共30个文件）
├── FM_broadcast_0.npy ... _29.npy （共30个文件）
├── ISM_sensors_0.npy ... _29.npy （共30个文件）
├── NOAA_APT_0.npy ... NOAA_APT_29.npy （共30个文件）
├── NOAA_weather_0.npy ... _29.npy （共30个文件）
├── noise_0.npy ... noise_29.npy （共30个文件）
└── pager_0.npy ... pager_29.npy （共30个文件）

每份.npy文件包含以下内容：
- **数据形状：** (1048576,) —— 对应1秒时长、1.024 MSPS采样率的信号
- **数据类型：** `complex64`（同相/正交（I/Q）采样数据）
- **单文件大小：** 约8.4 MB

## 采集硬件
- **软件定义无线电（SDR）设备：** RTL-SDR Blog V4（售价39.95美元）
- **主控计算机：** Indiedroid Nova 16GB版本（售价179.95美元）
- **天线：** 伸缩偶极天线（标配）

## 模型性能
当使用随机森林（100棵决策树）训练时：
- **整体准确率：** 87.5%
- **完美分类类别：** ADS-B、FM广播、ISM传感器、NOAA APT、NOAA气象、寻呼机（分类准确率100%）
- **易混淆类别：** APRS与噪声（信号包稀疏，易出现分类混淆）

## 引用格式
bibtex
@misc{rtl-ml-dataset,
author = {TrevTron},
title = {RTL-ML数据集：经过验证的射频信号采集集},
year = {2026},
publisher = {Hugging Face},
howpublished = {url{https://huggingface.co/datasets/TrevTron/rtl-ml-dataset}}
}

## 许可证
MIT许可证 —— 可免费用于商业与非商业用途。

## 相关资源
- **代码仓库：** [github.com/TrevTron/rtl-ml](https://github.com/TrevTron/rtl-ml)
- **个人博客：** [unland.dev](https://unland.dev)（即将上线）
- **硬件搭建指南：** [Indiedroid Nova 设备搭建](https://github.com/TrevTron/rtl-ml/blob/main/docs/HARDWARE_SETUP.md)

## 贡献说明
本数据集于美国加利福尼亚州特曼库拉（南加州）采集，采集条件如下：
- 与多个信号源保持清晰视距
- 通过频谱分析与人工检查完成样本验证
- 所有样本均满足最低信噪比要求（调制信号信噪比≥10 dB）

如有疑问或改进建议，请访问[GitHub仓库](https://github.com/TrevTron/rtl-ml)。