electricsheepafrica/radar-multimodal-edge

---
license: apache-2.0
task_categories:
- tabular-classification
language:
- en
tags:
- radar
- sensor-fusion
- multimodal
- imu
- acoustic
- edge-computing
- drone-detection
- vibration-analysis
- synthetic-data
pretty_name: "Multi-Modal Edge Sensor Signatures (Radar + IMU + Acoustic)"
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: data/features.parquet
---

# Multi-Modal Edge Sensor Signatures (Radar + IMU + Acoustic)

## Dataset Description

Synthetic multi-modal sensor fusion dataset combining radar, inertial measurement unit (IMU), and acoustic signatures for edge-deployed threat classification nodes. Contains 2,100 labeled samples across 7 target classes, with features extracted from all three modalities plus cross-modal coherence measures. Designed for scenarios where any single sensor modality may be degraded (RF jamming, acoustic noise, poor ground coupling).

### Dataset Summary

| Property | Value |
|----------|-------|
| **Total samples** | 2,100 (300 per class, perfectly balanced) |
| **Classes** | 7 |
| **Features** | 44 columns (40 numeric, 4 categorical/ID) |
| **Sensors** | X-band radar + MEMS IMU (MPU6050) + acoustic microphone |
| **Range** | 100 m to 3 km |
| **SNR range** | 0 to 47 dB |
| **Format** | Apache Parquet |

## Sensor Specifications

| Sensor | Parameter | Value |
|--------|-----------|-------|
| **Radar** | Type | X-band phased array |
| | Frequency | 9.5 GHz |
| | PRF | 20 kHz |
| | CPI | 256 pulses |
| **IMU** | Model | MPU6050 (MEMS accelerometer) |
| | Sample rate | 1,000 Hz |
| | Range | +/- 16 g |
| | Noise density | 400 ug/sqrt(Hz) |
| | Axis | Z (vertical, ground vibration) |
| **Acoustic** | Type | Microphone |
| | Sample rate | 8,000 Hz |
| | Sensitivity | -38 dB |
| | Noise floor | 30 dB SPL |

## Target Classes

| Class | Group | Category | Radar RCS (m^2) | Key IMU Signature | Key Acoustic Signature |
|-------|-------|----------|-----------------|-------------------|------------------------|
| `drone_air` | 1 | Threat | 0.01 | 267/534 Hz rotor harmonics (0.002 g) | 267/534/801 Hz harmonics |
| `drone_ground` | 1 | Threat | 0.01 | 267/534 Hz coupled (0.05 g, near-field) | 267/534 Hz (louder, close) |
| `vehicle_ground` | 2 | Threat | 10.0 | 30/60/90/120 Hz engine (0.5 g, strong) | 80/160/240 Hz engine |
| `personnel` | 2 | Threat | 1.0 | 1.5/3.0 Hz gait (0.01 g, weak) | 200/400/800 Hz footsteps |
| `aircraft_distant` | 0 | Clutter | 50.0 | 0.5 Hz (barely detectable) | 100/200 Hz (faint) |
| `wind_clutter` | 0 | Clutter | distributed | 5/15/30 Hz broadband | 100/200/400 Hz broadband |
| `rain_clutter` | 0 | Clutter | distributed | 50/100/200 Hz impact | 1000/2000/4000 Hz |

## Multi-Modal Signal Generation

### Radar
Range-Doppler maps are generated with the same coherent pulse-Doppler pipeline as other datasets. Class-specific micro-Doppler (rotor, vibration, distributed) is injected as amplitude-modulated sidebands. Environmental effects (attenuation, multipath scintillation) are applied to the RD power map.

### IMU (Vibration)
Synthetic accelerometer signals are generated by summing sinusoidal components at class-specific vibration frequencies with:
- Harmonic content (2nd harmonics at 30% amplitude)
- Frequency jitter (+/- 5% variation per component)
- Amplitude variation (+/- 50% per component)
- MEMS noise floor (400 ug/sqrt(Hz))
- Range-dependent ground coupling attenuation: `1 / (1 + range/50m)`

### Acoustic
Acoustic signals are synthesized from class-specific frequency components with:
- SPL-calibrated amplitudes: `A = 10^(SPL/20) * 20 uPa`
- Frequency jitter (+/- 2%)
- Ambient noise floor (30 dB SPL)
- Inverse-distance attenuation

### Environmental Effects on Modalities
- **Wind/rain on acoustic**: Environments with high clutter spectral width add broadband noise
- **Urban on IMU**: High-multipath environments add ambient ground vibration
- **Rain on radar**: Rain reflectivity raises the effective noise floor

## Features

### Radar Features (15)
- `rd_peak_power_db`: Peak power in RD map (dB)
- `rd_peak_range_bin`: Range bin of peak
- `md_mean_doppler_hz`, `md_std_doppler_hz`: Doppler spectral center and spread
- `md_energy_spread`: Energy concentration in top signal bins
- `md_dominant_freq_hz`: Peak Doppler frequency
- `bf_flash_rate_hz`: Blade flash periodicity
- `bf_max_doppler_spread_hz`: Maximum Doppler spread from config
- `bf_modulation_index`: Sideband-to-main-lobe energy ratio
- `st_md_bandwidth_hz`: 90% energy bandwidth (peak-centered)
- `st_md_periodicity_hz`: Envelope FFT periodicity
- `st_md_contrast`: Peak-to-median power contrast
- `snr_db`: Signal-to-noise ratio (0-55 dB)
- `rcs_estimated_m2`, `rcs_true_m2`: Estimated and ground-truth RCS

### IMU Features (10)
- `imu_peak_freq_hz`: Dominant vibration frequency
- `imu_dominant_amplitude_g`: Peak spectral amplitude (g)
- `imu_spectral_centroid_hz`: Energy-weighted mean frequency
- `imu_bandwidth_hz`: 3-dB bandwidth
- `imu_rms_g`: RMS acceleration
- `imu_crest_factor`: Peak-to-RMS ratio
- `imu_num_peaks`: Number of spectral peaks above 10% of max
- `imu_low_freq_energy`: Fractional energy 0-50 Hz
- `imu_mid_freq_energy`: Fractional energy 50-200 Hz
- `imu_high_freq_energy`: Fractional energy 200-500 Hz

### Acoustic Features (9)
- `ac_peak_freq_hz`: Dominant acoustic frequency
- `ac_spl_db`: Sound pressure level (relative dB)
- `ac_spectral_centroid_hz`: Energy-weighted mean frequency
- `ac_bandwidth_hz`: 3-dB bandwidth
- `ac_harmonic_ratio`: Fraction of peaks at harmonic multiples
- `ac_num_harmonics`: Detected harmonic peaks
- `ac_low_freq_energy`: Fractional energy 0-500 Hz
- `ac_mid_freq_energy`: Fractional energy 500-2000 Hz
- `ac_high_freq_energy`: Fractional energy 2000-4000 Hz

### Cross-Modal Features (3)
- `xmodal_radar_imu_coherence`: Radar-IMU spectral correlation [-1, 1]
- `xmodal_radar_ac_coherence`: Radar-acoustic spectral correlation [-1, 1]
- `xmodal_imu_ac_coherence`: IMU-acoustic spectral correlation [-1, 1]

### Scenario Metadata (4)
- `range_m`: Target range (m)
- `velocity_ms`: Radial velocity (m/s)
- `environment`: Environment condition
- `jamming_type`: EW type (always "none" for edge nodes)

## Usage

```python
from datasets import load_dataset

ds = load_dataset("electricsheepafrica/radar-multimodal-edge")
df = ds['train'].to_pandas()

# Modality-specific feature groups
radar_features = [c for c in df.columns if c.startswith(('rd_', 'md_', 'bf_', 'st_', 'snr'))]
imu_features = [c for c in df.columns if c.startswith('imu_')]
acoustic_features = [c for c in df.columns if c.startswith('ac_')]
cross_modal = [c for c in df.columns if c.startswith('xmodal_')]

# Sensor fusion: use all modalities
all_features = radar_features + imu_features + acoustic_features + cross_modal + ['rcs_estimated_m2']
X = df[all_features].values
```

## Data Quality

- Perfectly balanced: 300 samples per class
- No NaN or Inf values
- RCS estimation correlates with ground truth (r ~ 0.94)
- Micro-Doppler features decoupled from peak power via peak-centered windowing
- Cross-modal coherence reflects genuine signal structure correlation, not artifacts
- Environment effects propagate realistically across all three modalities

## Citation

```bibtex
@misc{electricsheep2025multimodal,
title={Multi-Modal Edge Sensor Signatures: Radar, IMU, and Acoustic Fusion for Threat Classification},
author={Electric Sheep Africa},
year={2025},
publisher={HuggingFace},
url={https://huggingface.co/datasets/electricsheepafrica/radar-multimodal-edge}
}
```