SUDARSHANchakra - Acoustic UAV Threat Detection System

Named after the divine discus weapon of Lord Vishnu, this system serves as a protective guardian against aerial threats.

A production-grade deep learning system for detecting drone/UAV acoustic signatures using Mel-Spectrogram analysis and Convolutional Neural Networks.

Mission

Develop an acoustic surveillance system for ground-based defense units that can:

• Listen to environmental audio in real-time
• Detect the specific acoustic signature of drone propellers
• Classify audio as Threat (Drone) or Safe (Background)

Features

• Audio-to-Vision Pipeline: Converts raw audio waveforms to Mel-Spectrograms for CNN analysis
• Custom CNN Architecture: 4-layer network optimized for acoustic pattern recognition
• Defense-Grade Metrics: Prioritizes Recall (minimizing missed threats) over Precision
• Production Ready: Modular Python package with inference API
• Auto-Ingestion: Automatically downloads and prepares the DroneAudioDataset

Project Structure

SudarshanChakra/
├── configs/
│   └── config.py           # Central configuration parameters
├── src/
│   ├── __init__.py
│   ├── data_ingestion.py   # Auto-clone dataset from GitHub
│   ├── data_loader.py      # PyTorch Dataset & DataLoader
│   ├── model.py            # CNN architectures
│   ├── train.py            # Training pipeline with early stopping
│   └── inference.py        # Real-time threat detection
├── outputs/
│   ├── models/             # Saved model checkpoints
│   ├── plots/              # Confusion matrix, training curves
│   └── logs/               # Training reports (JSON)
├── data/                   # Auto-downloaded dataset
├── main.py                 # Main entry point
├── requirements.txt        # Dependencies
└── README.md

Quick Start

1. Setup Environment

# Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

2. Train the Model

# Complete pipeline: download data -> train -> evaluate
python main.py --train

This will:

1. Clone the DroneAudioDataset from GitHub
2. Print the dataset directory structure
3. Auto-discover Drone/Background audio paths
4. Train the CNN with early stopping
5. Generate confusion matrix and training plots
6. Save the best model to outputs/models/best_model.pth

3. Run Inference

# Analyze a single audio file
python main.py --detect path/to/audio.wav

# With custom threshold (lower = more sensitive)
python main.py --detect recording.wav --threshold 0.3

# Run demo on sample data
python main.py --demo

Configuration

Edit configs/config.py to modify system parameters:

Parameter	Default	Description
SAMPLE_RATE	22050 Hz	Audio sampling rate
DURATION	2.0 s	Analysis window
N_MELS	128	Mel frequency bins
BATCH_SIZE	32	Training batch size
LEARNING_RATE	1e-4	Adam optimizer LR
THREAT_CONFIDENCE_THRESHOLD	0.4	Detection sensitivity

Dataset

Source: DroneAudioDataset by Sara Al-Emadi

The system automatically:

1. Clones the repository to data/DroneAudioDataset/
2. Scans for Binary_Drone_Audio folder structure
3. Discovers Drone and Background/Unknown audio directories
4. Validates WAV file integrity

Model Architecture

DroneDetectorCNN (Default)

Input: Mel-Spectrogram (1, 128, 87)
    ↓
ConvBlock: 1 → 32 channels, MaxPool
    ↓
ConvBlock: 32 → 64 channels, MaxPool
    ↓
ConvBlock: 64 → 128 channels, MaxPool
    ↓
ConvBlock: 128 → 256 channels, MaxPool
    ↓
Global Average Pooling
    ↓
FC: 256 → 128 → 64 → 2 (with Dropout)
    ↓
Output: [Safe, Threat] logits

Total Parameters: ~500K (lightweight for edge deployment)

Defense Metrics

In defense applications:

• False Negative (missed drone) = Critical → Prioritize high Recall
• False Positive (false alarm) = Acceptable → Accept lower Precision

The system uses a 0.4 confidence threshold by default, biasing toward threat detection.

Output Example

[TEST RESULTS]
  Accuracy:  0.9523
  Precision: 0.9412
  Recall:    0.9697  ← 97% of drones detected!
  F1 Score:  0.9552

[DEFENSE METRICS INTERPRETATION]
  - Recall (96.97%): Percentage of actual threats detected
    → 3.0% of threats are MISSED (False Negatives)
  - Precision (94.12%): Percentage of alerts that are real threats
    → 5.9% of alerts are FALSE (False Positives)

Outputs

After training, find these artifacts in outputs/:

File	Description
models/best_model.pth	Best model checkpoint
plots/confusion_matrix.png	Test set confusion matrix
plots/training_history.png	Loss, accuracy, metrics curves
logs/training_report.json	Full training metrics log

Inference API

from src.inference import ThreatDetector

# Initialize detector
detector = ThreatDetector(threshold=0.4)

# Analyze audio file
result = detector.detect("recording.wav")

# Result format:
# {
#     "status": "THREAT",  # or "SAFE"
#     "confidence": 0.89,
#     "probabilities": {"safe": 0.11, "threat": 0.89},
#     "file": "recording.wav"
# }

# Print formatted alert
detector.print_alert(result)

CLI Reference

# Show configuration
python main.py --config

# Data ingestion only
python main.py --ingest

# Training pipeline
python main.py --train

# Inference
python main.py --detect audio.wav
python main.py --detect audio.wav --threshold 0.3

# Demo mode
python main.py --demo

Requirements

• Python 3.8+
• PyTorch 2.0+
• librosa 0.10+
• scikit-learn 1.3+
• Git (for dataset cloning)

Citation

If using the DroneAudioDataset:

Al-Emadi, Sara, et al. "Audio Based Drone Detection and Identification
using Deep Learning." 2019 15th International Wireless Communications
& Mobile Computing Conference (IWCMC). IEEE, 2019.

License

This project is for authorized defense research and educational purposes.

---

SUDARSHANchakra - Protecting the skies through acoustic intelligence.